Lifelong Multi-Agent Path Finding Based on Reinforcement Learning and GPU

Abstract

Lifelong Multi-Agent Path Finding (LMAPF) is vital for logistics and production as automation scales. Existing real-time methods rely on distributed, learning-based frameworks but often depend on search-based single-agent planners, creating time bottlenecks. They also use complex communication structures fixed to agent counts, limiting the scalability and time efficiency. We propose RL-LMAPF, a scalable real-time framework that introduces two key innovations. First, the Multi-target GPU-based Breadth-First Search (MGBFS) algorithm extends traditional GPU-accelerated BFS to compute cost maps for multiple targets simultaneously, significantly reducing memory consumption and computation time. Second, we introduce a computationally efficient communication module that is independent of agent numbers. Our design significantly reduces computational overhead, which prioritizes critical inter-agent interactions while ensuring scalability for massive agent counts. Experiments show RL-LMAPF handles up to tens of thousands of agents with decision times around 1 second per step.