Multi-agent Pathfinding with Local and Global Guidance

Multi-agent path finding (MAPF) exists in many practical applications, such as intelligent warehouses. In this type of scenario, the agents need to cooperate with each other and eventually reach the target point without collision. The existing multi-agent path planning algorithms are mainly centralized algorithms, such as Conflict-Based Search (CBS). However, this kind of approach is difficult to solve the problem in real time and its scalability is poor. In this work, we focus on solving MAPF problem in intelligent warehouse. To address this problem, instead of using time-consuming search based algorithms, we propose a novel decentralized multi-agent pathfinding method based on deep reinforcement learning. Combined with curriculum learning, the algorithm uses local and global guidance mechanisms to help agents plan feasible paths. As a result, the success rate of the algorithm has been significantly improved. Experimental results show that our algorithm generalizes well and it still performs well when the scale of problem increases. The solution efficiency is close to the centralized algorithms.