efficient informative sensing using multiple robots amarjeet singh, andreas krause, carlos guestrin...

Download Efficient Informative Sensing using Multiple Robots Amarjeet Singh, Andreas Krause, Carlos Guestrin and William J. Kaiser (Presented by Arvind Pereira

If you can't read please download the document

Post on 19-Dec-2015

216 views

Category:

Documents


2 download

TRANSCRIPT

  • Slide 1
  • Efficient Informative Sensing using Multiple Robots Amarjeet Singh, Andreas Krause, Carlos Guestrin and William J. Kaiser (Presented by Arvind Pereira for CS-599 Sequential Decision Making in Robotics)
  • Slide 2
  • 2 Predicting spatial phenomena in large environments Constraint: Limited fuel for making observations Fundamental Problem: Where should we observe to maximize the collected information? Biomass in lakes Salt concentration in rivers
  • Slide 3
  • 3 Challenges for informative path planning Use robots to monitor environment Not just select best k locations A for given F(A). Need to take into account cost of traveling between locations cope with environments that change over time need to efficiently coordinate multiple agents Want to scale to very large problems and have guarantees
  • Slide 4
  • 4 How to quantify collected information? Mutual Information (MI): reduction in uncertainty (entropy) at unobserved locations [Caselton & Zidek, 1984] MI = 4 Path length = 10 MI = 10 Path length = 40
  • Slide 5
  • 5 Y1Y1 Y2Y2 Y3Y3 Y4Y4 Y5Y5 Selection B = {Y 1,, Y 5 } Key observation: Diminishing returns Y1Y1 Y2Y2 Selection A = {Y 1, Y 2 } Adding Y will help a lot! Adding Y doesnt help much Y New observation Y Y B A + + Large improvement Small improvement For A B, F(A [ {Y}) F(A) F(B [ {Y}) F(B) Submodularity: Many sensing quality functions are submodular*: Information gain [Krause & Guestrin 05] Expected Mean Squared Error [Das & Kempe 08] Detection time / likelihood [Krause et al. 08] *See paper for details
  • Slide 6
  • 6 Selecting the sensing locations Lake Boundary G1G1 G2G2 G3G3 G4G4 Greedy selection of sampling locations is (1-1/e) ~ 63% optimal [Guestrin et. al, ICML05] Result due to Submodularity of MI: Diminishing returns Greedy may lead to longer paths! Greedily select the locations that provide the most amount of information
  • Slide 7
  • 7 Greedy - reward/cost maximization Available Budget = B s Reward = B Cost = B reward cost = 2 reward cost = 1
  • Slide 8
  • 8 Greedy - reward/cost maximization Available Budget = B- s B B B Too far! Greedy Reward = 2
  • Slide 9
  • 9 Greedy - reward/cost maximization Available Budget = 0 s B B Greedy Reward = 2 Optimal Reward = B Greedy can be arbitrarily poor!
  • Slide 10
  • 10 Informative path planning problem max p MI(P) MI submodular function Lake Boundary Start- s Finish- t P C(P) B Informative path planning special case of Submodular Orienteering Best known approximation algorithm Recursive path planning algorithm [ Chekuri et. Al, FOCS05]
  • Slide 11
  • 11 Recursive path planning algorithm [Chekuri et.al, FOCS05] Start (s) Finish (t) vmvm Recursively search middle node v m P1P1 P2P2 Solve for smaller subproblems P 1 and P 2
  • Slide 12
  • 12 v m2 Recursive path planning algorithm [Chekuri et.al, FOCS05] Start (s) Finish (t) P1P1 v m1 v m3 Maximum reward Recursively search v m C(P 1 ) B 1 Lake boundary vmvm
  • Slide 13
  • 13 Recursive path planning algorithm [Chekuri et.al, FOCS05] Start (s) Finish (t) P1P1 vmvm Recursively search v m C(P 1 ) B 1 Commit to the nodes visited in P 1 Recursively optimize P 2 C(P 2 ) B-B 1 P2P2 Maximum reward Committing to nodes in P 1 before optimizing P 2 makes the algorithm greedy!
  • Slide 14
  • 14 Quasi-polynomial running time O (B*M) log(B*M) B: Budget Reward Chekuri Reward Optimal log(M) M: Total number of nodes in the graph 6080100120140160 Cost of output path (meters) 0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000 Execution Time (Seconds) OOPS ! Small problem with 23 sensing locations Recursive path planning algorithm [Chekuri et.al, FOCS05]
  • Slide 15
  • 15 Almost a day!! Recursive path planning algorithm [Chekuri et.al, FOCS05] Quasi-polynomial running time O (B*M) log(B* M) B: Budget Reward Chekuri Reward Optimal log(M) M: Total number of nodes in the graph Small problem with 23 sensing locations
  • Slide 16
  • Recursive-Greedy Algorithm (RG)
  • Slide 17
  • 17 Selecting sensing locations Given: finite set V of locations Want: A * V such that Typically NP-hard! Greedy algorithm: Start with A = ; For i = 1 to k s* := argmax s F(A [ {s}) A := A [ {s*} G1G1 G2G2 G3G3 G4G4 How well does the greedy algorithm do?
  • Slide 18
  • 18 Selecting sensing locations Given: finite set V of locations Want: A * V such that Typically NP-hard! Greedy algorithm: Start with A = ; For i = 1 to k s* := argmax s F(A [ {s}) A := A [ {s*} G1G1 G2G2 G3G3 G4G4 Theorem [Nemhauser et al. 78] : F(A G ) (1-1/e) F(OPT) Greedy near-optimal!
  • Slide 19
  • Sequential Allocation
  • Slide 20
  • Sequential Allocation Example
  • Slide 21
  • Spatial Decomposition in recursive-eSIP
  • Slide 22
  • recursive-eSIP Algorithm
  • Slide 23
  • SD-MIPP
  • Slide 24
  • eMIP
  • Slide 25
  • Branch and Bound eSIP
  • Slide 26
  • Experimental Results
  • Slide 27
  • Experimental Results : Merced
  • Slide 28
  • Comparison of eMIP and RG
  • Slide 29
  • Comparison of Linear and Exponential Budget Splits
  • Slide 30
  • Computation Effort w.r.t Grid size for Spatial Decomposition
  • Slide 31
  • Collected Reward for Multiple Robots with same starting location
  • Slide 32
  • Collected Reward for Multiple Robots with different start locations
  • Slide 33
  • Paths selected using MIPP
  • Slide 34
  • Running Time Analysis Worst-case running time for eSIP for linearly spaced splits is: Worst-case running time for eSIP for exponentially spaced splits is: Recall that Recursive Greedy had:
  • Slide 35
  • Approximation guarantee on Optimality
  • Slide 36
  • Conclusions eSIP builds on RG to near-optimally solve max collected information with upper bound on path-cost SD-MIPP allows multiple robot paths to be planned while providing a provably strong approximation gurantee Preserves RG approx gurantee while overcoming computational intractability through SD and branch & bound techniques Did extensive experimental evaluations