Reinforcement Learning for Large-Scale Lifelong Multi-Agent Path Finding with Shared Spatial Fields

Key points

• Large-scale applications of lifelong multi-agent path finding (LMAPF) call for computationally inexpensive movement decisions under changing traffic conditions and across varied layouts.1
• The complete policy is trained end-to-end using average-reward reinforcement learning with throughput as the reward, without planner demonstrations or a separate online path planner.1
• For structured and unstructured layouts, we train one policy per class on mixtures of small maps and deploy each without retraining on larger, unseen layouts and agent populations.1
• Across six layouts with 10000 agents, our policies achieve throughput comparable to or higher than the strongest evaluated baselines.1

Sources (1)

• [1]Reinforcement Learning for Large-Scale Lifelong Multi-Agent Path Finding with Shared Spatial Fields
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 13, 11:25 AM

  We propose a navigation policy in which a convolutional network computes a shared spatial field over the map once per timestep and a compact goal-conditioned operator derives agents' movement preferences.

  Large-scale applications of lifelong multi-agent path finding (LMAPF) call for computationally inexpensive movement decisions under changing traffic conditions and across varied layouts.

Extractive summary: sentences quoted from the sources.