Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Cooperative Framework for Multiple Moving Agents

Cooperative Framework for Multiple Moving Agents


  1. Cooperative Framework for Multiple Moving Agents Keisuke Okumura School of

    Computing, Tokyo Institute of Technology Nov. 26th, 2019
  2. /22 2 Moving AIs YouTube/Mind Blowing Videos Twitter/@PDChina Twitter/@knaohiro1 Systems

    with moving agents are becoming common. YouTube/NISSAN Twitter/@intel Twitter/@Elecrow1 required: cooperation/coordination What are? How to achieve? Generic structure?
  3. /22 3 Case Study - Exploration [Okumura et al., iCAST-18]

    building a map of unknown environment with multi-robot Who should visit? process 2 Find paths for multiple agents process 3 Where should be visited? process 1 known unknown limited sensing ability frontier
  4. /22 4 Case Study - Warehouse [Ma et al.,AAMAS-17] conveying

    of packages in lifelong situations Who should visit? process 2 Find paths for multiple agents process 3 Where should be visited? process 1 delivery loc. pickup loc. given (Multi-agent Pickup & Delivery, MAPD)
  5. /22 Generic Model given - environment (graph) - agents (with

    locations) collaboration process 1. task generation specify target nodes 2. task allocation decide which agent visits where 3. path finding give valid paths basis /22 5
  6. /22 6 given agents (starts) graph goals obtain 1. paths

    without collisions s.t. all agents are on their goals simultaneously In real scenarios, MAPF must be solved online and iteratively Multi-agent Path Finding (MAPF)
  7. /22 7 Iterative MAPF given B C A agents (starts)

    graph tasks (set of goals) X obtain 1. paths without collisions s.t. all tasks are completed in finite time 2. task allocation A C A B C [Okumura et al., IJCAI-19]
  8. /22 8 Design Choice of Algorithm optimality vs speed tradeoff

    globally locally vs communication Adaptivity for Iterative Use centralized decentralized vs style (relatively)
  9. /22 9 Priority Inheritance with Backtracking for Iterative Multi-agent Path

    Finding Keisuke Okumura, Manao Machida, Xavier Défago, Yasumasa Tamura IJCAI-19 ü propose an algorithm PIBT for solving iterative MAPF while ensuring reachability in specific graphs, e.g., biconnected 500 agents within 0.5 sec all agents reach their goals in finite time after being given
  10. /22 10 Priority Inversion high low mid STUCK node =

    resource priority inversion in resource scheduling prioritized planning s.t. only determines the next locations [Silver, AIIDE-05]
  11. /22 11 Priority Inheritance countermeasure to priority inversion [Sha et

    al., 1990] priority inheritance high mid low as high
  12. /22 12 Stuck Again Priority inheritance is insufficient to prevent

    all stuck situations. high as high as high as high as high STUCK
  13. /22 13 with Backtracking high Agents with priority inheritance have

    to wait for backtracking. invalid valid valid You can move invalid You must re-plan, I will stay re-plan re-plan
  14. /22 14 Guarantee of Local Movement The agent with highest

    priority successfully moves to a target node if the two nodes (current, target) belong to a simple cycle. Lemma. local movement highest current target move to an arbitrary neighbor node move along with the shortest path Giving dynamic priorities, any agent that has not reached goal, gets the highest eventually!
  15. /22 15 Reachability (PIBT) If a graph has a cycle

    for all pairs of adjacent nodes, then, with PIBT, all agents reach their own goals within timesteps after being given. Theorem. reachability G <latexit sha1_base64="yVMSfsXKkBgam/Vt6xvyf13//ws=">AAACTXicbZFLSwMxFIUz9V2f1aWbYBFclRkVdVl0oUsFq4JTyp30tgbzGJKMWob6B9zq73LtD3EnYqYWrNYLgcN3ciD3JEkFty4M34LSxOTU9MzsXHl+YXFpeaWyemF1Zhg2mBbaXCVgUXCFDcedwKvUIMhE4GVye1T4l3doLNfq3PVSbEroKt7hDJxHZ8etlWpYCwdDx0U0FFUynNNWJdiL25plEpVjAqy9jsLUNXMwjjOB/XKcWUyB3UIXr71UINE288FL+3TTkzbtaOOPcnRARxM5SGt7MvE3Jbgb+9cr4I+n8J5pKUG18xhMV8JDP48LT6d5bCT17LGAseCSO9sfT3D1T8LDn8ToNq5z0My5SjOHin0v08kEdZoW1dI2N8ic6HkBzHDfB2U3YIA5/wFl33T0t9dxcbFdi3Zq4dlutX447HyWrJMNskUisk/q5ISckgZhBMkTeSYvwWvwHnwEn99XS8Ews0Z+TWnmC2o3tic=</latexit> diam(G) · |A| <latexit sha1_base64="Asn8QhFjVGk7W4BNG/bJ+1yoTx8=">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</latexit> #agents NOT ensure that all agents be on their goals simultaneously! biconnected directed ring not biconnected
  16. /22 16 Exp.1 Classical MAPF Parallel Push & Swap [Sajid

    et al., SoCS-12] [Silver, AIIDE-05] WHCA*-10 PIBT ost003d, 194x194 [Sturtevant, 2012] 300 agents *average over only success cases within the solver (in 100 times) path cost (steps) #agents computation time (sec.) #agents success (%) #agents
  17. /22 17 Exp.2 Multi-agent Pickup & Delivery [Ma et al.,

    AAMAS-17] PIBT TP [Ma et al., AAMAS-17] task = (pickup loc., delivery loc.) let free agents move to the nearest pickup loc. PIBT with task allocation 50 agents, 500 tasks service time (steps) (task) frequency (task) frequency comp. time (sec)
  18. /22 18 Multi-agent Pickup & Delivery Sushi

  19. /22 Extensions / Variants 19 today ongoing PIBT [Okumura et

    al., IJCAI-19] future direction relax graph cond. generalize time-window be async with dynamic time-window adaptive winPIBT realize! winPIBT [arXiv:1905.10149] Causal PIBT in arbitrary graph Generic PIBT+TI PIBT+TI 1.Dest. 2. Path Finding 3.Instruction 4. Move 5.Verify AFADA
  20. /22 20 winPIBT: Extended Prioritized Algorithm for Iterative Multi-agent Path

    Finding Keisuke Okumura, Yasumasa Tamura, Xavier Défago arXiv:1905.10149 ü propose a generalization of PIBT with respect to time-window, called winPIBT high low ideal sum of path cost winPIBT-50 : PIBT = 0.85 : 1
  21. /22 21 Prioritized Planning with Temporary Inflation for Iterative Multi-agent

    Path Finding Keisuke Okumura, Yasumasa Tamura, Xavier Défago presented in FIT-19 ü aiming to expand target graphs, introduce temporary inflation mechanism to PIBT (PIBT+TI) root root upstream downstream be high be Low PIBT PIBT+TI
  22. /22 22 Conclusion cooperative framework 1. task generation 2. task

    allocation 3. path finding model iterative MAPF algorithm PIBT, winPIBT, PIBT+TI Magic Roundabout Wikipedia 1000 agents provide adaptive pathfinding method for countless agents REALITY communication time-assumption crash/fault TOY MODEL self-interested … …