cooperative motion and task planning under …introduction nominal scenario recon guration...

98
Introduction Nominal Scenario Reconfiguration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic Control Lab, EES Royal Institute of Technology, KTH, Sweden Licentiate Seminar Meng Guo Licentiate Seminar 1

Upload: others

Post on 28-Jun-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Cooperative Motion and Task PlanningUnder Temporal Tasks

Meng Guo

Automatic Control Lab EESRoyal Institute of Technology KTH Sweden

Licentiate Seminar

Meng Guo Licentiate Seminar 1

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 2

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

MotivationIndustrial and domestic robots1 boosted by

I Development of digital processing unitsbull more powerful in speed and capacity

bull more affordable

I Wireless communication technologybull potentially connects the robots

bull integrates with other ldquosmartrdquo devices

1Gostai Roomba Romo Amigo service robotsMeng Guo Licentiate Seminar 3

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

MotivationIndustrial and domestic robots1 boosted by

I Development of digital processing unitsbull more powerful in speed and capacity

bull more affordable

I Wireless communication technologybull potentially connects the robots

bull integrates with other ldquosmartrdquo devices

1Gostai Roomba Romo Amigo service robotsMeng Guo Licentiate Seminar 3

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-users

bull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planning

bull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 2: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 2

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

MotivationIndustrial and domestic robots1 boosted by

I Development of digital processing unitsbull more powerful in speed and capacity

bull more affordable

I Wireless communication technologybull potentially connects the robots

bull integrates with other ldquosmartrdquo devices

1Gostai Roomba Romo Amigo service robotsMeng Guo Licentiate Seminar 3

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

MotivationIndustrial and domestic robots1 boosted by

I Development of digital processing unitsbull more powerful in speed and capacity

bull more affordable

I Wireless communication technologybull potentially connects the robots

bull integrates with other ldquosmartrdquo devices

1Gostai Roomba Romo Amigo service robotsMeng Guo Licentiate Seminar 3

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-users

bull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planning

bull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 3: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

MotivationIndustrial and domestic robots1 boosted by

I Development of digital processing unitsbull more powerful in speed and capacity

bull more affordable

I Wireless communication technologybull potentially connects the robots

bull integrates with other ldquosmartrdquo devices

1Gostai Roomba Romo Amigo service robotsMeng Guo Licentiate Seminar 3

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

MotivationIndustrial and domestic robots1 boosted by

I Development of digital processing unitsbull more powerful in speed and capacity

bull more affordable

I Wireless communication technologybull potentially connects the robots

bull integrates with other ldquosmartrdquo devices

1Gostai Roomba Romo Amigo service robotsMeng Guo Licentiate Seminar 3

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-users

bull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planning

bull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 4: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

MotivationIndustrial and domestic robots1 boosted by

I Development of digital processing unitsbull more powerful in speed and capacity

bull more affordable

I Wireless communication technologybull potentially connects the robots

bull integrates with other ldquosmartrdquo devices

1Gostai Roomba Romo Amigo service robotsMeng Guo Licentiate Seminar 3

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-users

bull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planning

bull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 5: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-users

bull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planning

bull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 6: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-users

bull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planning

bull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 7: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-users

bull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planning

bull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 8: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motivation

I Imagine several care robots in your house

bull ldquoAmigo go to the kitchen find anapple and bring it to merdquo

bull ldquoGostai keep an eye on the kids inthe living room and bedroomrdquo

bull ldquoAmigo put all the toys in thebasketrdquo

bull middot middot middot

Meng Guo Licentiate Seminar 4

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-users

bull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planning

bull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 9: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-users

bull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planning

bull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 10: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots that

bull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planning

bull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 11: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planning

bull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 12: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Challenges

Challenges for system engineers (us)

I Provide the non-expert end-usersbull a formal but flexible way to specify daily tasks

bull task execution status as feedback

I Design algorithms for robots thatbull synthesize and execute a plan to satisfy the task

bull without or with minimal human intervention

bull accommodate changes in the workspace

bull initiate communication with other devices

bull handle collaborative tasks

Meng Guo Licentiate Seminar 5

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planning

bull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 13: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planning

bull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 14: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 15: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis

2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 16: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Background

I Motion and task planningbull motion plan of dynamic systems2

bull task plan for discrete-event systems3

I Model checkingbull for verification4

bull for plan synthesis2S M LaValle Planning algorithms 20063M Ghallab et al Automated planning theory amp practice 20044C Baier J-P Katoen Principles of model checking 2008

Meng Guo Licentiate Seminar 6

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 17: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Multi-agent System

I Multi-agent system to solve a global task

bull decompose into sub-tasks

bull top-down tightly-coupled

I Multi-agent system with local tasks

bull favour individual interests

bull bottom-up loosely-coupled

Meng Guo Licentiate Seminar 7

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 18: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Main Contributions

I Reconfiguration for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Motion and Action

Planning under LTL Specification using Navigation Functions andAction Description Language IROS Japan 2013

bull M Guo and D V Dimarogonas Reconfiguration in Motion Planningof Single- and Multi-agent Systems under Infeasible Local LTLSpecifications CDC Italy 2013

I Real-time adaptation for single- and multi-agent systemsbull M Guo K H Johansson and D V Dimarogonas Revising Motion

Planning under Linear Temporal Logic Specifications in PartiallyKnown Workspaces ICRA Germany 2013

bull M Guo and D V Dimarogonas Distributed Plan Reconfiguration viaKnowledge Transfer in Multi-agent Systems under Local LTLSpecifications ICRA Hongkong 2014

bull M Guo and D V Dimarogonas Multi-agent Plan Reconfigurationunder Local LTL Specifications IJRR Submitted

Meng Guo Licentiate Seminar 8

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 19: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 9

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 20: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW

bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 21: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion Abstraction

I Abstracted as the weighted finite transition system (FTS)

Tc = (Π minusrarrc Π0 AP Lc Wc)

where the finite regions Π = π1 π2 middot middot middot πW bull properties of interest AP = APr capAPpbull Π0 initial states

bull minusrarrcsube ΠtimesΠ control-driven transition

bull Lc Π rarr 2AP labelling function

bull Wc ΠtimesΠ rarr R+ weight function

I Examples

1

2 3

4

Meng Guo Licentiate Seminar 10

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 22: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Abstraction Outcome

Meng Guo Licentiate Seminar 11

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 23: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 24: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Temporal Task Specification

How to formally specify the task

I Language Linear-time Temporal Logic (LTL) formula

I Syntax

ϕ = True | a | ϕ1 or ϕ2 | notϕ | copy ϕ | ϕ1 Uϕ2

I Semantics5

I Specified over AP

I To specify control tasks

bull Safety notϕ1 Order diams(ϕ1 and diams(ϕ2 and diamsϕ3))

bull Response ϕ1 rArr ϕ2 Liveness diamsϕ1

5C Baier J-P Katoen Principles of model checking 2008Meng Guo Licentiate Seminar 12

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 25: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)

I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 26: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 27: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Task Interpretation

I ldquogo to the kitchen find anapple and bring it to merdquo

I ldquokeep an eye on the kids inthe living room andbedroomrdquo

I ldquoput all the toys in thebasketrdquo

I middot middot middot

I ϕ =diams((Kitand PickApp)anddiams(Liv)

)I ϕ = diamsLiv anddiamsBed

I ϕ = diams

(PickToyrarr(

notPickToy U (Bas and

DropToy)))

I middot middot middot

Meng Guo Licentiate Seminar 13

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 28: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 29: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 30: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

Given the FTS Tc and the LTL task ϕ

I Synthesize a discrete plan that satisfies ϕ

I Construct the hybrid control strategy to execute thederived plan

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

0 500 1000 1500 20000

500

1000

1500

2000

Meng Guo Licentiate Seminar 14

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 31: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Nominal Solution Outline

I Automata-based model-checking algorithm6

I Hybrid controller synthesis7

6C Baier J-P Katoen Principles of model checking 20087G E Fainekos et al Temporal logic motion planning for dynamic

robots Automatica 2009Meng Guo Licentiate Seminar 15

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 32: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 33: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 1 Translation

I Translate ϕ into the Nondeterministic Buchi automaton(NBA) Aϕ over 2AP

Aϕ = (Q 2AP δ Q0 F)

q1

q2

q3

q1q1 nota4

a1 amp nota4

nota4

a2 amp nota4

nota4

a1 amp nota4

a2 amp a3 amp nota4

nota4a1 amp a2 amp a3 amp nota4

a3 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp nota4

a1 amp a2 amp a3 amp nota4

q0

bull Q is a finite set of states

bull 2AP are alphabets

bull δ sube Qtimes 2AP timesQ

bull Q0 F are initial accepting states

bull χ(qm qn) = l isin 2AP | (qm l qn) isin δ

I Automated

I Fast translation algorithm8

8D Oddoux P Gastin LTL2BA software 2009Meng Guo Licentiate Seminar 16

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 34: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 35: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 2 Product Automaton

I Construct the weighted product automaton Ap = Tc otimesAϕ

Ap = (Qprime δprime Qprime0 F prime Wp)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesFbull δprime sube QtimesQ where (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc

and (qm Lc(πi) qn) isin δ

bull Wp δprime rarr R+ Wp((〈πi qm〉 〈πj qn〉)) = Wc(πi πj)

I Algorithms

bull static construction

bull on-the-fly construction

Meng Guo Licentiate Seminar 17

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 36: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Structure and Cost

I Accepting run R of Ap with the prefix-suffix structure

R = Rpre (Rsuf)ω = qprime0 q

prime1 middot middot middot qprimefminus1

[qprimef q

primef+1 middot middot middot qprimen

I The total cost

Cost(R Ap) =

fminus1sumi=0

Wp(qprimei qprimei+1) + γ

nminus1sumi=f

Wp(qprimei qprimei+1)

Meng Guo Licentiate Seminar 18

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 37: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 38: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 3 Graph Search

I Algorithm based on Nested-Dijkstra shortest path

bull shortest path from every qprime0 isin Qprime0 to every qprimef isin F prime

bull shortest cycle from qprimef and backP_FFP_IF

I The derived accepting run Ropt with minimal cost

I Corresponding plan τ = Ropt|Π

I Complexity O(|δprime| middot log2 |Qprime| middot (|Qprime0|+ |F prime|))

Meng Guo Licentiate Seminar 19

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 39: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Step 4 Hybrid Control Strategy

To execute τ = Ropt|Π using U(x(t) πi πj)

I generate an infinite execution with finite representation

I monitor the execution status

I record past motions

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

Meng Guo Licentiate Seminar 20

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 40: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 41: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Shortcomings of Nominal Solution

I Reconfiguration

bull plan as sequence of motion (no actions)

bull feasible task

I Real-time adaptation

bull fully-known workspace

bull plan synthesized once off-line

bull executed regardless of the real observation

I Multi-agent system with local tasks

bull independent or dependent tasks

bull communication

bull collaborative tasks

Meng Guo Licentiate Seminar 21

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 42: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 22

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 43: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Motion and Action Planning

Motion Planning

Geometric Constrains

Workspace Model

Model-checking

TaskAction Planning

Discrete Planning

PropositionalLogic

I Why is it necessarybull to automate the choice of actions

bull plan as sequence of motion and actions

I Why plan motion and action togetherbull the purpose of ldquogoing somewhererdquo is to ldquodo somethingrdquo

I Why model motion and action separatelybull robotrsquos mobility depend on the workspace structure

bull robotrsquos capable actions relatively fixed

Meng Guo Licentiate Seminar 23

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 44: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Mobility and Action

I Mobility abstraction is given by (similarly as Tc)

M = (ΠM actM minusrarrM ΠM0 ΨM LM WM)

I Capable action set ActB = actB0 actB1 middot middot middot actBKI Precondition function

Cond ActB times 2Ψp times 2Ψs minusrarr TrueFalse

I Effect function

Eff ActB times (2Ψs timesΨb) minusrarr (2Ψs timesΨb)

I Mimic action description language (ADL)

Meng Guo Licentiate Seminar 24

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 45: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Model of Complete Functionalities

hellip

|||

M

Bhellip

R

I Complete functionalities by composing M and B

R = (ΠR ActR minusrarrR ΠR0 ΨR LR WR)

I Automated algorithm

Meng Guo Licentiate Seminar 25

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 46: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 47: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Much richer task specifications

I Given ϕ over robot motion action and internal states

I Accepting run with prefix-suffix structure with similardefinition for cost

I Plan as a sequence of motion and action minimal cost

I Construct the hybrid controller accordingly

r1 c1 c2 c3 r3

c3 c2 r2 c2 r5

c2 r2 c2 c1(r1)ω

r1 c1 c2 c3 r3PickG

c3 c2 r2Drop c2 r5PickR

c2 r2Drop c2 c1(r1)ω

Meng Guo Licentiate Seminar 26

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 48: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 49: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goal

bull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 50: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Potentially Infeasible Task

I Def 26 ϕ is infeasible for Tc if no accepting run of Ap canbe found

I Closely related to partially-known workspace

I Examplebull ldquogo to the kitchen find an apple and bring it to merdquo

rArr infeasible if no apple

bull ldquokeep an eye on the kids in the living room and bedroomrdquorArr infeasible if not known where the bedroom is

I Nominal solution fails

I Our goalbull synthesize the plan satisfying the task partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 27

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 51: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 52: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I Relaxed product automaton Ar = TctimesAϕ

Ar = (Qprime 2AP δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime sube Qprime timesQprime (〈πi qm〉 〈πj qn〉) isin δprime iff (πi πj) isinminusrarrc andexist l isin 2AP such that (qm l qn) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)= Wc(πi πj) + α middot Dist(L(πi) χ(qm qn))

I α ge 0 user-defined balance

Meng Guo Licentiate Seminar 28

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 53: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Balanced Cost

I Accepting run R with prefix-suffix structure

R = qprime0 qprime1 middot middot middot [ qprimef qprimef+1 middot middot middot qprimen ]ω

= 〈π0 q0〉 〈π1 q1〉 middot middot middot [ 〈πf qf 〉 middot middot middot 〈πn qn〉 ]ω

I The balanced cost of an accepting run

Cost(R Ar) =

fminus1sumi=0

Wr(qprimei qprimei+1) + γ

nminus1sumi=f

Wr(qprimei qprimei+1)

= costτ + α middot distϕ

I costτ implementation cost distϕ distance to ϕ

I distϕ = 0rArr R|Π |= ϕ

Meng Guo Licentiate Seminar 29

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 54: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|Π

I Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 55: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I The balanced accepting run

Rbal = minR

Cost(R Ar)

I The balanced plan τbal = Rbal|ΠI Proposed algorithms

bull synthesize Rbal given γ and α

bull computes the associated costτ and distϕ for τbal

I Theorem 31 If ϕ is feasible over Tc the balanced plan τbal

satisfies ϕ if α gt α

Meng Guo Licentiate Seminar 30

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 56: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 57: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Feedback by Tuning αI α tunable balance between costτ and distϕ

bull α uarr rArr distϕ darr rArr satisfy ϕ more

bull α darr rArr costτ darr rArr cheaper

I Example ϕ = diamsa1 andnot(a2 and a3)

π1

π2

π3

a2a3

a1

a2

30

40 20

10Φ

π0 a1qq1

a1

not a2 amp not a3 not a2 amp not a3

q05 5

0 5 10 15 20 25 3030

40

50

60

70

80

90

100

110

120Cost of the Optimal Accepting Path

To

tal C

ost

by

(7)

30 40 50 60 70 80 900

1

2

3

4

5

6

7

8

Implementation Cost

Dis

tan

ce t

o S

pec

ific

atio

n

Potential Motion Plans

Plan1 (0)

Plan2 0

1(

3)

Plan3 0

2(

3)

Meng Guo Licentiate Seminar 31

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 58: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 59: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goal

bull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 60: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Soft and Hard Specification

I Specification with two distinctive parts

ϕ = ϕhard and ϕsoft

I ϕhard for safety or securitybull ldquodo not go to the balconyrdquo ldquoalways alarm if see firerdquo

I ϕsoft for additional achievement (maybe infeasible)bull ldquocollect the toys in all roomsrdquo

I Nominal solution fails

I Our goalbull synthesize the plan satisfies ϕhard completely and ϕsoft

partially

bull user-defined balance on satisfiability and cost of the plan

Meng Guo Licentiate Seminar 32

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 61: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 62: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

SolutionI Relaxed intersection of Asoft and Ahard as Aϕ by Alg 9

Aϕ = (Q 2AP δ Q0 F)

I Safety-ensured and relaxed product automaton

Ar = TctimesAϕ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠtimesQ Qprime0 = Π0 timesQ0 F prime = ΠtimesF

bull δprime Qprime rarr 2Qprime 〈πj qn〉 isin δprime(〈πi qm〉) iff (πi πj) isinminusrarrc and

qn isin δ(qm Lc(πi))

bull Wr δprime rarr R+ is the weight function

Wr(〈πi qm〉 〈πj qn〉)=Wc(πi πj) + α middot Dist(Lc(πi) χsoft(q2 q2))

Meng Guo Licentiate Seminar 33

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 63: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 64: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Theorem 33 Assume R is an accepting run of Arτ = R|Π is safe for Tc and ϕ

I Balanced cost of accepting runs

Cost(R Ar) = costτ + α middot distϕsoft

I The safe accepting run

Rsafe = minR

Cost(R Ar)

I The safe plan τsafe = Rsafe|ΠI Similar algorithms to synthesize Rsafe

I Lemma 32 distϕsoft = 0rArr τsafe |= ϕ

Meng Guo Licentiate Seminar 34

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 65: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

On-line Planning

Why put planner on-line

I To handle partially-knownworkspace

I Plan may not be executedas expected

I Plan could be improved

I Feed real-time observationback to planner

Meng Guo Licentiate Seminar 35

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 66: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Problem Formulation

I The agentrsquos FTS at time t ge 0

T tc = (Π minusrarrtc Π0 AP L

tc W

tc )

I Task specification

ϕ = ϕsoft and ϕhard

I Our goal

bull model the robotrsquos sensing

bull update the system model

bull guarantee the plan is always valid and safe

Meng Guo Licentiate Seminar 36

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 67: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution Framework

Meng Guo Licentiate Seminar 37

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 68: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 69: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 70: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Initial Synthesis and System Update

I Step 1 initial synthesis at t = 0

bull τ0 obtained for feasible or infeasible task

bull starts executing τ0

I Step 2 knowledge update at t ge 0

bull sensing information obtained

Senset = ((π S Snot) E Enot)

I Step 2 update T tc based on Senset

Meng Guo Licentiate Seminar 38

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 71: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 72: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 73: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Plan Verification and Revision

I Step 3 validate the current planbull validity lArrrArr invalid transitions

bull safety lArrrArr unsafe transitions

I Step 4 local revision instead of full synthesis

I Low complexity suitable for real-time applications

I Theorem 35 validity and safety of the revised planguaranteed

Meng Guo Licentiate Seminar 39

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 74: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Simulation Example

I Nonholonomic ground vehicle

I Potential filed-based navigation controller9

I Surveillance over regions 2 3 4

I Detect walls and obstacles in real-time

1

23

5

6

7

9

10

11

12

8

4

1

2 3

4

9S R Lindemann I I Hussein S M LaValle Real time feedbackcontrol for nonholonomic mobile robots with obstacles CDC 2006

Meng Guo Licentiate Seminar 40

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 75: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 41

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 76: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 77: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependent Local Tasks

I System of N agents i = 1 middot middot middot N

Ti = (Πi minusrarri Πi0 APi Li Wi)

I Locally-assigned LTL specification by ϕi

I ϕi may contain requirements on other agents j 6= i

bull constraints eg ldquodo not be in the same room with agent 1rdquo

bull collaborations eg ldquomove the desk together with agent 1rdquo

I Difficultiesbull the joined execution may not be mutually feasible even

though the individual one is

bull the priority of each agent plays an important role

Meng Guo Licentiate Seminar 42

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 78: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 79: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Dependency Cluster

I Dependency Agents i and j are called dependent if

Θ3

Θ2

Θ1

Θ4

(1) agent i depends on agent j ifAPϕi

andAPj 6= empty or

(2) agent j depends on agent i ifAPϕj

andAPi 6= empty

I Dependency graph Gd = (V E) V = 1 2 middot middot middot N andE sube V times V is the dependency relation

I Dependency cluster Θ sube V foralli j isin Θ there is a path fromi to j in Gd

Meng Guo Licentiate Seminar 43

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 80: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 81: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Mutual Infeasible

Within one cluster Θ = 1 2 middot middot middot MI The composed FTS TΘ = T1 times middot middot middot TM is

TΘ = (ΠΘ minusrarrΘ ΠΘ0 APΘ LΘ WΘ)

I The mutual specification is

ϕΘ = ϕ1 and ϕ2 middot middot middot and ϕM

I Mutually infeasible if ϕΘ is infeasible over TΘ

I The relaxed intersection of Aϕi of ϕi i isin Θ

AϕΘ = (Q 2APϕΘ δ Q0 F)

Meng Guo Licentiate Seminar 44

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 82: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesF

bull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprimeΘ qb〉) isin δprime iff (πΘ π

primeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 83: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Solution

I The relaxed product automaton

ArΘ = TΘ times AϕΘ = (Qprime δprime Qprime0 F prime Wr)

where Qprime = ΠΘ timesQ Qprime0 = ΠΘ0 timesQ0 F prime = ΠΘ timesFbull δprime sube Qprime timesQprime (〈πΘ qa〉 〈πprime

Θ qb〉) isin δprime iff (πΘ πprimeΘ) isinminusrarrΘ

and (qa qb) isin δ

bull Wr δprime rarr R+ is the weight function

Wr(〈πΘ q1 middot middot middot qM t〉 〈πprimeΘ q

prime1 middot middot middot qprimeM tprime〉)

= WΘ(πΘ πprimeΘ) + α

Msumi=1

βi Dist(LΘ(πΘ) χi(qi qprimei))

I α penalty on violating ϕΘ

I βi priority of agent irsquo task

Meng Guo Licentiate Seminar 45

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 84: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 85: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Synthesize the balanced accepting run of ArΘI Projection onto Ti as the individual plan i isin Θ

I Change α and β

I Example

0 1 2 3 4 5 6 7 8 9 100

2

4

6

8

10

12

14

16

18

20

2 (

1=1)

Unique Optimal Accepting Paths

0 05 1 15 2 25 3 35 4 45 50

05

1

15

2

25

3

35

4

6

Distance to 1

Potential Motion Plans

10

2028

24

36

Dis

tan

ce t

o

2

Meng Guo Licentiate Seminar 46

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 86: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 87: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Independent Local Tasks

I System of N agents coexisting within a partially-knownworkspace

T ti = (Πi minusrarrti Πi0 APi L

ti W

ti )

I Locally-assigned task specification and independent

ϕi = ϕsofti and ϕhard

i

I Motivation

bull agents located at various locations within the workspace

bull observe up-to-date information

bull beneficial to communicate

Meng Guo Licentiate Seminar 47

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 88: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 89: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Knowledge Update and Transfer

I Knowledge update bybull own sensing ability Sensetk = (π S Snot) E Enotbull communication with others

I Communication network Nk isin N (static or dynamic)

I Transfer knowledge

bull request once Requesttkg = (k ϕk|APk)

bull event-based reply Replythk = (π Sprime Sprimenot) where

Sprime = S cap (ϕh|APh) and Sprime

not = Snot cap (ϕh|APh)

I Update T tk based on Sensetk and Replytgk

I Validate and revise the current plan

Meng Guo Licentiate Seminar 48

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 90: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Results

I Full synthesis or local revision

I Event-based trigger to re-synthesize the plan

0 20 40 60 80 100Time (s)

0

1000

2000

3000

4000

5000

6000

Tota

l Cos

t

[5522]

[3319]

[3132]

[139]

[139]

[5522]

[3319]

[1134]

[341]

[241]

[139]

[7512]

[0318]

[3130]

[139]

[6520]

[0227]

[0135]

[139]

[8520]

[2133]

[240]

[139]

Ar_1Ar_2Ar_3Ar_4Ar_5

0 10 20 30 40 50 60Time

0

2

4

6

8

10

12

14

o

f Mes

sage

s

Ar_1Ar_2Ar_3Ar_4Ar_5

Gw_1Gw_2Gw_3Gw_4Gw_5

Gf_1Gf_2Gf_3Gf_4Gf_5

Meng Guo Licentiate Seminar 49

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 91: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Software ImplementationI Robot operating system(ROS)-based

I ROS core + ROS nodes

I ROS node for planning

Meng Guo Licentiate Seminar 50

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 92: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Experiments

I NAO humanoid

I MAS Lab CVAP

I NEXUS ground vehicle

I Smart Mobility Lab ACL

Meng Guo Licentiate Seminar 51

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 93: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

IntroductionMotivationBackground

Nominal ScenarioProblem FormulationNominal Solution

Reconfiguration

Motion and ActionPotentially Infeasible TaskPartially-known Workspace

Multi-agentDependent Local TasksIndependent Local Tasks

SummarySummary

Meng Guo Licentiate Seminar 52

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 94: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 95: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 96: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Summary

I Motion and task planningbull Discrete motion and task plan with minimal costbull Hybrid control strategy

I Reconfiguration and real-time adaptationbull Potentially infeasible taskbull Soft and hard specificationsbull Partially-known workspacebull Motion and action planning

I Multi-agent systems with local tasksbull Dependent tasksbull Independent tasksbull Software implementation

Meng Guo Licentiate Seminar 53

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 97: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Future Work

I Automated abstraction

I Natural language to LTL graphic interface

I Trade-off between computational complexity andoptimality

I Robustness and fault tolerance (both motion and action)

I Continuous constraints coupled dynamics

Meng Guo Licentiate Seminar 54

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary
Page 98: Cooperative Motion and Task Planning Under …Introduction Nominal Scenario Recon guration Multi-agent Summary Cooperative Motion and Task Planning Under Temporal Tasks Meng Guo Automatic

Introduction Nominal Scenario Reconfiguration Multi-agent Summary

Thank you

Meng Guo Licentiate Seminar 55

  • Introduction
    • Motivation
    • Background
      • Nominal Scenario
        • Problem Formulation
        • Nominal Solution
          • Reconfiguration
            • Motion and Action
            • Potentially Infeasible Task
            • Partially-known Workspace
              • Multi-agent
                • Dependent Local Tasks
                • Independent Local Tasks
                  • Summary
                    • Summary