Slide 1

Slide 1 text

26th Jan. 2023 PhD Defense (Final) Planning, Execution, Representation, and Their Integration for Multiple Moving Agents Keisuke Okumura Tokyo Institute of Technology, Japan ౦ژ޻ۀେֶ 5PLZP*OTUJUVUFPG5FDIOPMPHZ http://bit.ly/3J96UjM

Slide 2

Slide 2 text

/56 2 YouTube/Mind Blowing Videos logistics YouTube/WIRED manufacturing YouTube/Tokyo 2020 entertainment Swarm { is, will be } necessary in everywhere

Slide 3

Slide 3 text

/56 3 objective-1 Representation objective-2 Planning Common Knowledge? Cooperation? (increased) Uncertainty Execution Navigation for a Team of Agents Who Plans? Huge Search Space

Slide 4

Slide 4 text

/56 4 integration Representation [AAMAS-22+] building effective roadmaps Execution [AAAI-21, ICRA-21, ICAPS-22*, IJCAI-22, AAAI-23+] overcoming uncertainties *Best Student Paper Award Planning ≥1000 agents within 1 sec [IJCAI-19, IROS-21, ICAPS-22*, AIJ-22, AAAI-23+] Research Summary

Slide 5

Slide 5 text

/56 5 Dissertation Overview

Slide 6

Slide 6 text

/56 6 Short-Horizon Planning for MAPF Short-Horizon Planning for Unlabeled-MAPF Short-Horizon Planning Guides Long-Horizon Planning Improving Solution Quality by Iterative Refinement 4. 5. 6. 7. Part I. Planning Building Representation from Learning Building Representation while Planning 11. 12. Part III. Representation Online Planning to Overcome Timing Uncertainties Offline Planning to Overcome Timing Uncertainties Offline Planning to Overcome Crash Faults 8. 9. 10. Part II. Execution Introduction, Preliminaries, and Background 1-3. Conclusion and Discussion 13. Dissertation Outline

Slide 7

Slide 7 text

/56 7 Short-Horizon Planning for MAPF Short-Horizon Planning Guides Long-Horizon Planning 4. 6. Part I. Planning Building Representation from Learning Building Representation while Planning 11. 12. Part III. Representation Recap: Quick Multi-Agent Path Planning discrete continuous

Slide 8

Slide 8 text

/56 8 Short-Horizon Planning for MAPF Short-Horizon Planning for Unlabeled-MAPF Short-Horizon Planning Guides Long-Horizon Planning Improving Solution Quality by Iterative Refinement 4. 5. 6. 7. Part I. Planning Dissertation Outline discrete

Slide 9

Slide 9 text

/56 9 MAPF: Multi-Agent Path Finding given agents (starts) graph goals solution paths without collisions optimization is intractable in various criteria [Yu+ AAAI-13, Ma+ AAAI-16, Banfi+ RA-L-17, Geft+ AAMAS-22]

Slide 10

Slide 10 text

/56 10 Challenge in Planning quality & solvability high low effort small large speed & scalability complete optimal incomplete suboptimal keeping solvability & solution quality with small planning effort holy grail

Slide 11

Slide 11 text

/56 11 Approaches in Planning quality & solvability effort high low small large speed & scalability approach-1 [Chap. 4-6] short-horizon planning guides long-horizon planning approach-2 [Chap. 7] iterative refinement planning time (sec) cost / lower bond 300 agents

Slide 12

Slide 12 text

/56 12 solved instances (%) runtime (sec) 0.0% A* [Hart+ 68] 0.4% ODrM* [W agner+ AIJ-15] 8.3% CBS [Sharon+ AIJ-15; Li+ AIJ-21] 10.7% BCP [Lam+ COR-22] 50.5% EECBS-5 [Li+ AAAI-21] 61.4% PP [Silver AIIDE-05] 80.9% LNS2 [Li+ AAAI-22] 67.4% PIBT [Okumura+ AIJ-22] complete & eventually optimal! 99.0% LaCAM* 13900 instances - 33 grid maps - random scenario - every 50 agents, up to max. (1000) tested on desktop PC blazing fast! SC & bounded sub-opt incomp & sub-opt SC* & optimal Results on MAPF Benchmark [Stern+ SOCS-19] *SC: solution complete: unable to distinguish unsolvable instances successfully breaking the trade-off comp & optimal 30.9% ODrM*-5 [W agner+ AIJ-15] comp & bounded sub-opt

Slide 13

Slide 13 text

/56 13 Building Representation from Learning Building Representation while Planning 11. 12. Part III. Representation Dissertation Outline continuous

Slide 14

Slide 14 text

/56 14 Background: Planning in Continuous Spaces given agents (starts) goals workspace necessity of constructing roadmap solution paths without collisions

Slide 15

Slide 15 text

/56 15 Challenge in Representation dense representation sparse representation produced by PRM [Kavraki+ 96] complete optimal incomplete suboptimal building small but promising representation for planning quality & solvability effort high low small large speed & scalability

Slide 16

Slide 16 text

/56 16 Approaches in Representation quality & solvability effort high low small large speed & scalability approach-1 [Chap. 11] learning from planning demonstrations approach-2 [Chap. 12] building representation while planning

Slide 17

Slide 17 text

/56 17 Future Direction integration LaCAM(*) PIBT planning in discretized spaces SSSP CTRMs planning in continuous spaces establish practical methodologies to multi-robot motion planning iterative refinement robust execution scalable & quick probabilistically complete asymptotically optimal, anytime with kinodynamic constraints

Slide 18

Slide 18 text

/56 18 Offline Planning to Overcome Timing Uncertainties Offline Planning to Overcome Crash Faults 9. 10. Part II. Execution Rest of Talk: Offline Planning with Reactive Execution

Slide 19

Slide 19 text

/56 19 Failure Demo of Synchronous Execution

Slide 20

Slide 20 text

/56 20 access:. 6th Jun. 2022 https://www.ft.com/content/aaddf4b1-a78b-4289-b42f-fd3f5cd7f176 “appears to have been caused by the collision of three bots on the grid” “more than 1,000 robots buzz around a grid, stopping to grab crates of food” Lacking robust execution may cause catastrophe

Slide 21

Slide 21 text

/56 21 Execution on Real Robots Planning Reality Gaps / Uncertainties frictions, sensor errors, communication delays, inconsistent speeds due to battery consumption, individual difference between robots, crash, etc overcoming uncertainties at runtime

Slide 22

Slide 22 text

/56 22 timing uncertainties crash faults

Slide 23

Slide 23 text

/56 23 timing uncertainties crash faults

Slide 24

Slide 24 text

/56 24 1 2 1 2 3 4 Planning Execution Delay Imperfect Execution

Slide 25

Slide 25 text

/56 25 1 2 1 2 3 4 Planning Execution Conservative: Synchronize Forcibly wait make agents wait for delayed ones arrival time: 3

Slide 26

Slide 26 text

/56 26 Progressive: Preserve Temporal Dependencies [Cap+ IROS-16, Ma+ AAAI-17] check temporal dependencies 1 2 1 2 3 4 => preserved, go arrival time: 2 1 2 1 2 3 4 Planning Execution 1 2 3 4 0 1 2 3 4 progress ideal actual homotopy

Slide 27

Slide 27 text

/56 27 Really Smart? 1 2 1 2 3 4 1 2 3 4 5 6 delay negative effect 60 agents, solved by PIBT Typical MAPF agents actions complicated dependencies [Okumura+ AIJ-22] [Okumura+ 22]

Slide 28

Slide 28 text

/56 28 Raspberry Pi x8 32 robots Bluetooth Expensive Communication Cost stable network / monitoring systems for ≥1000 agents…? NON-TRIVIAL!! [Okumura+ 22]

Slide 29

Slide 29 text

/56 29 online intervention at runtime to preserve timing-property abandon all timing assumptions

Slide 30

Slide 30 text

/56 30 Offline Time-Independent Multi-Agent Path Planning KO, François Bonnet, Yasumasa Tamura & Xavier Defago planning execution https://kei18.github.io/otimapp IJCAI-22 (extended version is under review at T-RO) formalize, analyze, and solve OTIMAPP OTIMAPP solution multi-agent pathfinding solution 2 1 4 3 0 0 1 2 weak to timing uncertainties

Slide 31

Slide 31 text

/56 31 given start goal graph solution path Problem Def. – OTIMAPP s.t. all agents eventually reach goals regardless of action orders

Slide 32

Slide 32 text

/56 32 Solution Analysis no reachable* cyclic deadlock two conditions are necessary & sufficient *non-reachable deadlocks exist no reachable* terminal deadlock

Slide 33

Slide 33 text

/56 33 Computational Complexity 1. finding solutions is NP-hard 2. verification is co-NP-complete main observations the proofs are reductions from 3-SAT OTIMAPP is computationally intractable

Slide 34

Slide 34 text

/56 34 Solvers both solvers can solve large OTIMAPP instances to some extent agents random-32-32-10 32x32 random-64-64-10 64x64 den520d 257x256 success rate (%) ≤ 5 min MAPF avoids collisions OTIMAPP avoids deadlocks propose two algorithms based on Multi-Agent Path Finding (MAPF) algorithms prioritized planning deadlock-based search extending conflict-based search [Sharon+ AIJ-15] extending conventional PP [Erdmann+ Algorithmica-87]

Slide 35

Slide 35 text

/56 35 Execution Demo no synchronization only local interactions centralized style with toio robots decentralized style with AFADA [Kameyama+ ICRA-21] all robots are guaranteed to reach their goals

Slide 36

Slide 36 text

/56 36 Conventional Solution Concept planning stage acting sage

Slide 37

Slide 37 text

/56 37 OTIMAPP Solution Concept eventually reaching goal status agents reactively switch executing plan

Slide 38

Slide 38 text

/56 38 timing uncertainties crash faults

Slide 39

Slide 39 text

/56 39 Fault-Tolerant Offline Multi-Agent Path Planning planning execution https://kei18.github.io/mappcf AAAI-23 fault solution multiple paths assuming crashes change execution path on-demand KO*, Sébastien Tixeuil formalize, analyze, and solve multi-agent path planning with crash faults (MAPPCF) *work done during staying in LIP6, Sorbonne University in France

Slide 40

Slide 40 text

/56 40 goal start graph 0 Conventional Solution Concept solution

Slide 41

Slide 41 text

/56 41 online replanning offline approach: preparing backup paths from the beginning or crash detected => then? 1 With Unforeseen Crash crashed (forever stop)

Slide 42

Slide 42 text

/56 42 primary path Solution Concept of MAPPCF backup path when is detected transition rule & 0 with crash faults

Slide 43

Slide 43 text

/56 43 Solution Concept of MAPPCF 1 with crash faults

Slide 44

Slide 44 text

/56 44 Solution Concept of MAPPCF 2 with crash faults

Slide 45

Slide 45 text

/56 45 Solution Concept of MAPPCF more than two agents may crash => backup path of backup path 3 done! with crash faults

Slide 46

Slide 46 text

/56 46 Problem Formulation of MAPPCF given solution s.t. all non-crashed agents eventually reach their destination, regardless of crashes (up to f ) & transition rules & maximum number of crashes f defined with failure detector & execution model centralized planning followed by decentralized execution

Slide 47

Slide 47 text

/56 47 Failure Detectors oracle that tells status of neighboring vertices response: 1. no agent query 2. non-crashed agent named FD 3. crashed agent anonymous FD unable to identify who crashes c.f., [Chandra+ JACM-96]

Slide 48

Slide 48 text

/56 48 Execution Models how agents are scheduled at runtime synchronous model all agents act simultaneously solutions avoid collisions MAPF: multi-agent pathfinding [Stern+ SOCS-19] solution solutions avoid deadlocks each agent acts spontaneously while locally avoiding collisions offline time-independent MAPP [Okumura+ IJCAI-22] sequential model (async) solution possible schedule

Slide 49

Slide 49 text

/56 49 Model Power Analyses SYN + AFD SYN + NFD SEQ + NFD SEQ + AFD synchronous model sequential model named failure detector anonymous FD SYN SEQ NFD AFD strictly stronger SEQ+AFD SYN+AFD solvable instances weakly stronger SYN+AFD SYN+NFD SYN+AFD SYN+NFD or solvable in SYN unsolvable in SEQ

Slide 50

Slide 50 text

/56 50 Computational Complexity 1. finding solutions is NP-hard 2. verification is co-NP-complete regardless of FD types or execution models the proofs are reductions from 3-SAT MAPPCF is computationally intractable

Slide 51

Slide 51 text

/56 51 Solver / Empirical Results MAPPCF provides better solution concept than finding disjoint paths random-32-32-10 32x32 (|V|=922) from [Stern+ SOCS-19] 30sec timeout with named FD success rate #agents DCRF/SYN disjoint paths fixed #crashes: f =1 adapted from CBS [Sharon+ AIJ-15] v.s. finding vertex disjoint paths costs / lower bound #agents DCRF/SYN disjoint paths traveling time when no crashes propose DCRF (decoupled crash faults resolution framework) to solve MAPPCF

Slide 52

Slide 52 text

/56 52 Summary in Execution planning execution reality gaps new solution concept: offline planning with reactive execution in practice? => hybrid of (relaxed) offline & online planning

Slide 53

Slide 53 text

/56 53 Closing

Slide 54

Slide 54 text

/56 TODAY 54 Planning Representation Execution integration [AAMAS-22+] building effective roadmaps [AAAI-21, ICRA-21, ICAPS-22, IJCAI-22, AAAI-23+] overcoming uncertainties ≥1000 agents within 1 sec [IJCAI-19, IROS-21, ICAPS-22, AIJ-22, AAAI-23+] Research Summary

Slide 55

Slide 55 text

/56 55 YouTube/Mind Blowing Videos warehouse YouTube/StarCraft video game [Flatland Challenge, AIcrowd] railway/airport operations [Morris+ AAAIW-16] [Le Goc+ UIST-16] robotic interface [Song+ ICCBB-01] drug discovery [Zhang+ 18] manufacturing [van den Berg+ ISRR-11] crowd simulation [Zhang+ SIGGRAPH-20] puzzle solving pipe routing [Belov+ SOCS-20] Impact of Research Results develop new horizons of multi-agent navigation contribution:

Slide 56

Slide 56 text

/56 More Info? => Check My Website! https://kei18.github.io/ Thank You for Listening!

Slide 57

Slide 57 text

/56 57 Publications 1. “Improving LaCAM for Scalable Eventually Optimal Multi-Agent Pathfinding.” KO (under review at IJCAI-23) 2. “Quick Multi-Robot Motion Planning by Combining Sampling and Search.” KO & Xavier Défago. (under review at IJCAI-23) 3. “LaCAM: Search-Based Algorithm for Quick Multi-Agent Pathfinding.” KO. AAAI. 2023. 4. “Fault-Tolerant Offline Multi-Agent Path Planning.” KO & Sebastien Tixeuil. AAAI. 2023. 5. “Priority Inheritance with Backtracking for Iterative Multi-agent Path Finding.” KO, Manao Machida, Xavier Défago & Yasumasa Tamura. Artificial Intelligence (AIJ). 2022 (previously presented at IJCAI-19) 6. “Offline Time-Independent Multi-Agent Path Planning.” KO, François Bonnet, Yasumasa Tamura & Xavier Défago. IJCAI. 2022. (extended version is under review at T-RO) 7. “Solving Simultaneous Target Assignment and Path Planning Efficiently with Time-Independent Execution.” KO & Xavier Défago. ICAPS. 2022. (best student paper award, extended version is under review at AIJ) 8. “CTRMs: Learning to Construct Cooperative Timed Roadmaps for Multi-agent Path Planning in Continuous Spaces.” KO, Ryo Yonetani, Mai Nishimura & Asako Kanezaki. AAMAS. 2022. 9. “Iterative Refinement for Real-Time Multi-Robot Path Planning.” KO, Yasumasa Tamura & Xavier Défago. IROS. 2021. Others: 10. “Active Modular Environment for Robot Navigation.” Shota Kameyama, KO, Yasumasa Tamura & Xavier Défago. ICRA. 2021. 11. “Time-Independent Planning for Multiple Moving Agents.” KO, Yasumasa Tamura & Xavier Défago. AAAI. 2021. 12. “Roadside-assisted Cooperative Planning using Future Path Sharing for Autonomous Driving.” Mai Hirata, Manabu Tsukada, KO, Yasumasa Tamura, Hideya Ochiai & Xavier Défago. VTC. 2021. 13. “winPIBT: Extended Prioritized Algorithm for Iterative Multi-agent Path Finding.” KO, Yasumasa Tamura & Xavier Défago. WoMAPF. 2021. 14. “Amoeba Exploration: Coordinated Exploration with Distributed Robots.” KO, Yasumasa Tamura & Xavier Défago. iCAST. 2018. under review peer-reviewed