Lifelong LaCAM with Local Guidance for Lifelong MAPF

Transcript

1. Lifelong LaCAM with Local Guidance for Lifelong MAPF Tomoki Arita , Keisuke Okumura 1,2 1 National Institute of Advanced Industrial Science and Technology (AIST), Japan 2 1 Keio University, Japan The 19th International Symposium on Combinatorial Search https://github.com/allegorywrite/lllg 1
2. Lifelong Multi Agent Path Finding (LMAPF) We study MAPF (multi agent path finding) used for warehouse automation Real warehouse require lifelong system Warehouse automation https://logistics.muratec.net/jp/products/premex/ 2
3. Lifelong Multi Agent Path Finding (LMAPF) Information (task, etc.) agent goal agent MAPF execution goal Arrived goals of all agents goal Throughput = timestep Planning action (1 step) Runtime = (per step) goal Runtime of planner timestep 3
4. Highlight slow / near optimal LLLG: Lifelong LaCAM with local guidance surpasses LMAPF performance frontier RHCR [Ma et al., AAAI-18] Lifelong LaCAM WPPL [Jiang et al., ICRA-25] LMAPF frontier PIBT [Okumura et al., SoCS-25] Guided PIBT [Chen et al., AAAI-24] Throughput Throughput LLLG (ours) fast / suboptimal Runtime [s / step] Runtime [s / step] Runtime [s / step] Runtime [s / step] 4
5. Lifelong MAPF is recursive one step execution horizon 1. search Call One shot MAPF 2. execution Call Call 5
6. (Tree) Search over Configurations for MAPF Configuration Configuration Generator (e.g. PIBT [Okumura et al., AIJ-22]. ) Feasible step t t+1 t+2 t+3 6
7. Lazy Constraints Addition search for MAPF (LaCAM) [Okumura, AAAI-23] PIBT Not immediately generated t PIBT t+1 t+2 PIBT t+3 7
8. Lazy Constraints Addition search for MAPF (LaCAM) [Okumura, AAAI-23] Adding constraints of agent’s move PIBT PIBT PIBT Not immediately generated t t+1 t+2 t+3 8
9. Guidance Myopic search is week for congestion → seeking guidance that can mitigate congestion Global guidance Local guidance (e.g. Guided PIBT [Chen et al., AAAI-24]) (e.g. LG-LaCAM [Arita et al., AAAI-26]) Guidance from start to goal Guidance for limited horizon Planning Guidance only at initial step Replanning at each step 9
10. Local Guidance PIBT without local guidance PIBT with local guidance 2nd 1st 1st 2nd 1. Create collision table (by guidance of other agents) repeat 2. Construct guidance by Space-time A* 3. Guidance biases Priorities of PIBT congestion ! 10
11. Revisit: (Tree) Search over Configurations for MAPF Construct guidance Construct guidance Construct guidance Construct guidance PIBT Not immediately generated t PIBT t+1 t+2 PIBT t+3 11
12. Revisit: Lifelong MAPF is recursive one step execution horizon 1. search Call One shot MAPF 2. execution Call Call 12
13. Lifelong MAPF is recursive one step execution windowed planning 1. search Call finite horizon full horizon 2. execution Call Call 13
14. Lifelong MAPF is recursive one step execution windowed planning 10 20 30 Call finite depth search (by LaCAM) 2. execution Call Call Throughput 1. search 3 Horizon=1 runtime 14
15. Lifelong MAPF is recursive one step execution horizon 1. search Call Utilize previous planning path 2. execution Call Call 15
16. Warm start guidance by previous planning path Previous planning path warm start Search Planned path Create collision table by planning path repeat Search Guidance Construct guidance by Space-time A* Lifelong 16
17. Warm start guidance by previous planning path Warm start guidance by previous guidance Guidance Search Search Planned path Search Guidance Search Lifelong Guidance Lifelong 17
18. Warm start guidance by previous planning path Search warm start by guidance Throughput Search warm start by previous path Planned path Guidance w/ inheritance RHCR runtime Lifelong random-64-64-20, 1000 agents 18
19. Heatmap Evaluation RHCR PIBT Throughput = 3.07 Throughput = 19.4 Guided PIBT LLLG (Ours) Throughput = 27.3 Throughput = 20.0 Throughput = 27.3 19
20. Heatmap Evaluation RHCR PIBT Guided PIBT LLLG Throughput = 3.07 Throughput = 19.4 Throughput = 20.0 Throughput = 27.3 stop counts of agents 7.3 + Global guidance + LaCAM + Local guidance 20
21. Heatmap Evaluation PIBT Lifelong LaCAM LLLG Throughput = 14.7 Throughput = 17.1 Throughput = 21.1 + LaCAM + LaCAM + Hindrance + Hindrance [Okumura et al., SoCS-25] + Local guidance stop counts of agents 21
22. Evaluation empty 48-48 LLLG (ours) warehouse 20-40-10-2-1 Throughput random 64-64-20 Runtime Runtime [s/step] Throughput LaCAM Guided PIBT PIBT RHCR WPPL den312d Number of agent 22
23. Summary github page We studies how to bring local guidance horizon 1. search Call Warm start by previous planning path 2. execution from MAPF to LMAPF The challenge is how to bring information from previous planning to current planning Call Call Our algorithm brings information Previous planning path warm starts of local guidance 23