Local Guidance for Configuration-Based Multi Agent Pathfinding

Transcript

1. Local Guidance for Configuration-Based Multi Agent Pathfinding Tomoki Arita , Keisuke Okumura 1,2 1 National Institute of Advanced Industrial Science and Technology (AIST), Japan https://github.com/allegorywrite/lg_lacam 2 1,3 Keio University, Japan 3 University of Cambridge, UK 1
2. Introduction Fleet management in factories is a critical task in robotics Multi-Agent Path Finding (MAPF) is an essential problem formulation for this task Warehouse automation Large-scale systems require fast and scalable algorithms https://logistics.muratec.net/jp/products/premex/ 2
3. Multi Agent Path Finding (MAPF) Trade-off in MAPF Slow but optimal Solution quality Fast and High quality algorithm Trade off Frontier LaCAM Fast but suboptimal Planning speed 3
4. Configuration LaCAM [Okumura, AAAI-23] (lazy constraints addition search for MAPF) Lazy successor generation by constraints addition with BFS (Breath First Search) Extremely fast and scalable algorithm (400 agents in 0.05 seconds) 4
5. LaCAM [Okumura, AAAI-23] (lazy constraints addition search for MAPF) Congested Area It's extremely fast but sensitive to congestion LaCAM 1000 800 600 400 Solution cost Obstacle map with 1K agent 200 Agents LNS2 [Li+ AAAI’22] (baseline) Runtime [s] 5
6. LG-LaCAM Introduce Local Guidance for MAPF ( like MPC … ) seeking guidance that can mitigate congestion priority ① ② ① ② ② ① PIBT Guided PIBT 6
7. Local Guidance Construction Initial guidance with collision (= congestion) Update agent 1 guidance Update agent 2 guidance Collision count: 5 Collision count: 2 Collision count: 1 construct local guidance sequentially for each agent 7
8. GG (Global Guidance) LG (Local Guidance) [Chen+ AAAI’24] Construct Local Guidance Construct Global Guidance PIBT PIBT … … replan PIBT … PIBT replan … Guidance from start to goal Guidance for limited horizon Planning Guidance only at initial step Replanning at each step Consider spatial Congestion Consider Spatio-Temporal Congestion 8
9. LG (Local Guidance) Achieve both congestion mitigation and performance improvement Smoothing out congestion, but still degrading performance. GG Solution cost Solution cost Obstacle map with 1K agent GG (Global Guidance) LNS2 (baseline) LG LNS2 (baseline) 1000 800 Runtime [s] 600 Runtime [s] 400 200 Agents 9
10. Obstacle map with 1K agent GG (Global Guidance) Fast but w/congestion LNS2 (baseline) Runtime [s] Solution cost Solution cost LaCAM Relaxing congestion GG Runtime [s] LG (Local Guidance) Solution cost Original LaCAM Resolving congestion further LG Runtime [s] 10
11. Huge Warehouse with 10K agent GG LaCAM LaCAM Runtime [s] LaCAM Solution cost Solution cost LNS2 LG (Local Guidance) (Global Guidance) Solution cost LaCAM Runtime [s] LG can handle ~10K agents, and outperforms both LNS2 and LaCAM Runtime [s] 11
12. Room map with 1K agent GG (Global Guidance) LG (Local Guidance) Bottleneck exists Smooth out bottleneck 12
13. Heatmap Visualization Room map LaCAM GG-LaCAM LG-LaCAM with 400 agents visit counts of agents Bottleneck exists Cannot deal with the bottleneck Smooth out bottleneck congestion 13
14. Various Maps Benchmark Significant cost reduction with reasonable runtime 14
15. All Maps Benchmark 644 scenarios × maps scenarios × maps 644644 senarios ×maps 15
16. Summary • Local guidance significantly improves configuration-based MAPF (e.g. LaCAM) Local Guidance Global Guidance Solution quality LaCAM New Frontier • Consistently outperforms global guidance and LNS-based refinements on many benchmarks. • Establishes a new Pareto frontier for large-scale, real-time MAPF Planning speed 16