Heuristic Large-Scale Multi-Agent Path Planning with Communication

Multi-agent path planning is an important topic in robotics and has significant implications in autonomous vehicles and intelligent logistics. In this paper, a heuristic multi-agent path planning method is proposed. The improved channel and spatial attention modules are utilized to extract key features from the maps of agents. Then communication is conducted through the spatio-temporal Graph Transformer, enhancing coordination among agents. To improve training efficiency, the heuristic path generated by expert algorithm is designed for model training. In addition, a distance-based function is applied to assist agents in escaping deadlock situations. Experimental results demonstrate that the proposed method performs well in multi-agent path planning tasks across various maps.