dynamic task allocation using multi-agent mobile robots · raghavv goel1, sha yi2, jaskaran singh...
TRANSCRIPT
Dynamic Task Allocation Using Multi-Agent Mobile RobotsRaghavv Goel1, Sha Yi2, Jaskaran Singh Grover2, Katia Sycara2
1 IIIT Delhi, 2 Carnegie Mellon University
Introduction
•Multi-agent systems are useful in surveillance,search and rescue, monitoring of crops etc.
•Above real world tasks are dynamic: constantlychanging and uncertain and we hence prefer usingmultiple agents as they provide provide wide-areacoverage, fault tolerance, and robustness tomissions.
•We address this issue of covering all the taskspresent in the environment by proposing analgorithm which solves an optimization problemto assign agents to different tasks after every timestep.
•Tasks are in the form of rooms as shown belowand as the agents only discover rooms when theyare in the doorways of the rooms, the proposedoptimization is solved at every time step.
Environment
Room 1
Room 8
Room 2
Room 7
Room 3
Room 6
Room 4
Room 5
Corridor
Leaf Room
Door
DoorDoor
Door
Door
Door
Door Door
Figure 1: These rooms are tasks present in the environment(which need to be completed), the room color denotes thetype of agent it needs and the darker a room is the morenumber of agents is required.
Figure 2: Environment without leaf room and same rooms,used for testing our proposed method.
Scenarios
Room type Agents Type Room req. Leaf Rooms1. same color same color fixed NA2. diff color RGB fixed NA3. diff color RGB varying A4. diff color RGB + white varying A
Method
We define a set of rooms(R), agents(A), doors,goals(G), available rooms(RA), checked rooms(RC),etc. and utility(U) and distance(D) matrix to solvean optimization problem based on constraints asshown below.
Scenario 1Solve below objective after every time step
maximizex λ1uᵀx + λ2d
ᵀx
subject to E1x = 1N ,TE2x ≥ TRreq,
E3x = 0M+1,
x ∈ BN(M+1)×1
(1)
where, λ1 ∈ [0, 1] and λ2 ∈ [0, 1],Scenario 2 (different color rooms & agents)Another constraint added to make same coloragent go to same room by using S ∈ B(M+1)×N
defined as
S(j, i) =
1 if color(Rj) = color(ai)0 otherwise
vec(S)ᵀx == N( no. of agents) (2)
Scenario 3 (Leaf room as shown in figure 1)1 Green leaf room undiscovered till blue agent enters blueroom up to certain point.
2 Agents made to move slowly when crossing differentcolored rooms, so that if same color room discovered, thenagents do not need to move in backward direction.
3 N adjacency matrix to account for the neighboursN ∈ B(M+1)×(M+1) defined as
N(j, i) =
1 if door present0 otherwise
Scenario 4 (White agents)1 White agents can enter any color room and then changecolor once room is checked.
2 Priority given to white agents.3 if a room has a neighbouring leaf room, then white agentscan reduce the number of other agents needed.
Results
0 20 40 60 80 100 120
Time(s)
2
4
6
8
10
Room
sR
oom
s
Room 5
Room 4
Room 3
Room 2
Room 1
Room 0
Times Steps
(a) Room available0 20 40 60 80 100 120 140 160
Time(s)
2
4
6
8
10
Room
sR
oom
s
Room 5
Room 4
Room 3
Room 2
Room 1
Room 0
Times Steps
(b) Room available
0 20 40 60 80 100 120
Time(s)
0
2
4
6
8
10
Room
sR
oom
s
Times Steps
Room 5
Room 4
Room 3
Room 2
Room 1
Room 0
(a) Rooms Checked0 20 40 60 80 100 120 140 160
Time(s)
0
2
4
6
8
10
Room
s
Room 2
Room 1
Room 0
(b) Rooms Checked
2
4
3
4
5
2.5
5.0
7.5
6
8
6
8
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0
x coordinate
0
5
10
y c
oo
rdin
ate
x coordinate
y c
oord
inate
Agent 1
Agent 2
Agent 3
Agent 4
Agent 5
Agent 6
(a) Agents trajectory
2.5
5.0
7.5
2.5
5.0
7.5
2.5
5.0
7.5
2.5
5.0
7.5
2.5
5.0
7.5
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0
x coordinate
0
5
10
y c
oo
rdin
ate
y c
oord
inate
x coordinate
Agent 1
Agent 2
Agent 3
Agent 4
Agent 5
Agent 6
(b) Agents trajectory(a) have λ1 = 1.0, λ2 = 0.0 while (b) have λ1 = 0.0, λ2 = 1.0.The avg time for checking all rooms in (a) = 132.84 while in(b) = 162.68 when ran over 25 trials.
0 20 40 60 80 100 120
2
4
6
8
10
Room
s
Times Steps
Room 5
Room 4
Room 3
Room 2
Room 1
Room 0
Rooms Available0 20 40 60 80 100 120
0
2
4
6
8
10
Room 5
Room 4
Room 3
Room 2
Room 1
Room 0
Rooms Checked
2
4
2
4
2.5
5.0
7.5
2.5
5.0
7.5
6
8
0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0
0
5
10
Agent 6
x coordinate
y c
oord
inate
Agent 1
Agent 2
Agent 3
Agent 4
Agent 5
Agents trajectory
Environment setting is as shown in figure 2 with the changesthat R1, R3 : green; R2, R5 : blue; R4 : red and A1, A2 :green; A3, A4 : blue; A5, A6 : red. The plots for the followingcases was the same:λ1 = 1.0 (b): λ2 = 0.0 and λ1 = 0.0, λ2 = 1.0. Hence, onlyplots for first case shown. Avg time for both was 136.
Conclusion and Discussion
•We observe that higher dependence on utility ui.e. λ1 closer to 1, leads to faster coverage versusdistance.
•However, when we introduce room colorconstraint then both the cases have almost equalmaximum time taken to check all the rooms andaverage checking time is same.
•We still need to integrate our latest approachwith barrier certificate to prevent inter agentcollisions. Also, we need to further build for lasttwo scenarios.
Future work
• Imagine a map with one room which has apyramid of neighbouring room and theseneighbours have more neighbours and so on, heresending as team of white agents seems the onlyfeasible solution.
•Learning based techniques were also exploredwith similar inclination in the form of predatorprey capturing, where the preys are like mobiletasks. But the environment was obstacle free andhence modelling an environment with obstaclesand restrictions on certain areas being allowed foragents needs to be done.
References
[1]A. J. Smith, G. Best, J. Yu, and G. A. Hollinger,“Real-time distributed non-myopic task selectionfor heterogeneous robotic teams,” 2019.
[2]A. Prorok, M. A. Hsieh, and V. Kumar, “Fastredistribution of a swarm of heterogeneousrobots,” 2016.
Acknowledgements
1 would also like to acknowledge the support of the RoboticsInstitute Summer Scholars program and its organizers: Dr.John Dolan and Ms. Rachel Burcin. Also 1 is really gratefulto all members of the AART Lab especially, Dr. Katia Sycara,Jaskaran Singh Grover, Sha Yi & Sumit Kumar