Real World Autonomous Agents Coordinate multiple agents Robust - PDF document

Massachusetts Institute of Technology Real World Autonomous Agents • Coordinate multiple agents Robust Execution of Contingent, • Provide robustness Temporally Flexible Plans Stephen Block Andreas Wehowsky, Brian Williams 2 Robustness to Disturbances Robustness to Temporal Uncertainty Robustness to ... • Temporal uncertainty : Temporally flexible mission plan Robustness to ... [l,u] • Temporal uncertainty : Temporally flexible mission plan S E Temporally flexible plans allow activities with uncertain duration [7,10] S E • Execution uncertainty : Dispatchable execution 7 , [ 0 ] 1 S E “Drive to rock” Simple temporal constraint : l ≤ t + - t - ≤ u T = t - T = t + • Communication latency : Distributed architecture [l,u] S E Drive to Rock Hardware command • Plan failure : Contingent mission plan 3 4 Robustness to Execution Uncertainty Robustness to Execution Uncertainty Robustness to ... Robustness to ... • Temporal uncertainty : Temporally flexible mission plan • Temporal uncertainty : Temporally flexible mission plan • Execution uncertainty : Dispatchable execution • Execution uncertainty : Dispatchable execution Temporally flexible plan Temporally flexible plan Assignment of Maintenance of Problem : Plan is either … Planning time Planning time execution times temporal flexibility • Brittle to temporal execution uncertainty Problem : Plan is either … • Overly conservative to ensure success • Brittle to temporal execution uncertainty • Overly conservative to ensure success Inflexible plan Temporally flexible plan Solution : Dispatchable execution … • Postpone scheduling until execution time Execution time Execution time Plan execution Reactively schedule execution times Hardware Commands Hardware Commands Hardware Hardware 5 6 1

Robustness to Execution Uncertainty Robustness to Execution Uncertainty Robustness to ... Robustness to ... • Temporal uncertainty : Temporally flexible mission plan • Temporal uncertainty : Temporally flexible mission plan • Execution uncertainty : Dispatchable execution • Execution uncertainty : Dispatchable execution Temporally flexible plan T=0 T=7 T=10 “Drive to rock” Maintenance of Problem : Plan is either … Planning time temporal flexibility • Brittle to temporal execution uncertainty [7,10] • Overly conservative to ensure success S E Drive to Rock Temporally flexible plan Solution : Dispatchable execution … [7,10] • Postpone scheduling until execution time S E Reformulation Execution time Drive to Rock Dispatchable plan Requires two-stage execution … Dispatching Least commitment planning allows the executive to use temporal • Plan reformulation to compile the plan to a form flexibility to respond to uncertainties at run time for easy dispatching Hardware Commands • Dispatching to schedule and execute activities Hardware 7 8 Robustness to Communication Robustness to Communication Latency Latency Robustness to ... Robustness to ... • Temporal uncertainty : Temporally flexible mission plan • Temporal uncertainty : Temporally flexible mission plan • Execution uncertainty : Dispatchable execution • Execution uncertainty : Dispatchable execution • Communication latency : Distributed architecture • Communication latency : Distributed architecture Problem : Problem : Centralized architecture Centralized architecture introduces communication introduces communication bottleneck at master agent bottleneck at master agent Solution : A distributed architecture evens out the communication requirements 9 10 Distributed Architecture Distributed Architecture Contingent temporally flexible plan Plan distribution Planning time Temporally flexible plan Execution time Reduced computational complexity Reformulation Dispatchable plan Avoids communication bottleneck Dispatching Hardware Commands Hardware 11 12 2

Plan Distribution Plan Distribution Exploit common hierarchical structure p1 if network is a single activity processor takes activity else if processor has followers p2 p3 distribute subnetworks to followers and self else if processor has co-leaders distribute subnetworks to coleaders and self leader election else p4 p5 processor takes entire network 0 ] [ 0 0 , , 0 [ ] [ 0 0 ] , , 0 ] [ 0 [ 0 [0,0] , 0 ] 13 14 Robustness to Plan Failure Distributed Architecture Contingent temporally Robustness to ... flexible plan • Temporal uncertainty : Temporally flexible mission plan • Execution uncertainty : Dispatchable execution Plan distribution • Communication latency : Distributed architecture • Plan failure : Contingent mission plan Reduced computational complexity Plan selection Planning time Temporally flexible plan Execution time Reduced computational complexity Reformulation Dispatchable plan Avoids communication bottleneck Dispatching Hardware Commands Hardware 15 16 Interleaved Candidate Generation and Candidate Generation Consistency Checking 1. Generate candidate plans through 1. Generate candidate plans through distributed search on the TPN distributed search on the TPN Interleaved and concurrent Interleaved and concurrent 2. Test the generated plans for 2. Test the generated plans for temporal consistency temporal consistency Implemented using a message passing scheme … Implemented using a message passing scheme … findfirst Initial search for a consistent set of choice variable assignments findfirst Initial search for a consistent set of choice variable assignments findnext Search for a new consistent assignment, to achieve global consistency findnext Search for a new consistent assignment, to achieve global consistency No consistent set of choice variable assignments was found No consistent set of choice variable assignments was found fail fail A consistent set of choice variable assignments was found A consistent set of choice variable assignments was found ack ack 17 18 3

Candidate Generation Consistency Checking Nodes send findfirst messages to their children Search progresses at increasing depth 1. Generate candidate plans through distributed search on the TPN Interleaved and concurrent 2. Test the generated plans for temporal consistency Implemented with synchronized distributed Bellman Ford Single Source Shortest Path algorithm • Requires only local knowledge • Uses a message passing scheme • Runs on distance graph corresponding to active portions of the TPN • Processor synchronization at every update cycle gives linear run time findfirst 19 20 Consistency Checking Candidate Generation Nodes perform consistency checking and report to their parent Nodes send findfirst messages to their children Checking progresses at decreasing depth Search progresses at increasing depth findfirst 21 22 Consistency Checking Time Complexity Analysis Nodes perform consistency checking and report to their parent Checking progresses at decreasing depth Centralized Planner DTP Candidate Generation N e N e Temporal Consistency Checking N 2 logN + NM N Overall Time Complexity Exponential Exponential N Number of nodes M Number of edges e Size of domain of choice variables • Worst case complexity of candidate generation corresponds to a plan entirely composed of choice nodes • DTP uses parallel processing for parallel and sequence networks • Time complexity much lower in realistic examples • In practice, complexity of DTP is near-linear 23 24 4

Questions ? 25 5