PRIMAL3 is a learning-based framework for ultra-large-scale multi-agent pathfinding. It combines reinforcement learning, imitation learning, topology-aware communication, LaCAM3-guided training, and PIBT-based action refinement. Agents use features related to cut vertices, dead ends, shortest paths, and blocking. Separate graphs model compatible following and conflicting movement directions. The paper reports scalability to city-level instances with up to 100,000 agents, along with ablations and physical-robot experiments. These claims come from the supplied abstract; detailed datasets, baselines, metrics, and experimental protocols require verification in the full paper.
No heat snapshots are available in the last 24 hours.