Multi-Agent Path Finding with Edge-Enhanced Graph Attention Network

With the widespread use of autonomous driving and automated storage, the problem of multi-robot path finding has become a popular research domain. Compared with centralized approaches, the advantage of decentralized approaches is that they avoid a sharp increase in computational costs as the number of robots increases, making them more suitable for large-scale problems. Efficient communication is paramount to the success of decentralized methods, and determining which information is beneficial for the task and which neighbor's information deserves more attention are key challenges that require careful consideration and resolution. However, existing methods are often too simplistic in aggregating information from neighboring robots, which limits the ability of robots to distinguish the importance of neighbor information in some scenarios. To address this issue, we propose a novel combination model that consists primarily of a Convolutional Neural Network (CNN) and an enhanced Graph Attention Network (GAT). The CNN is responsible for extracting features from the local environment of the robots. In order to improve the robot's ability to distinguish between important and irrelevant information, the edge features and node features are combined to compute the graph attention. We refer to this improved network as the Edge-Enhanced Graph Attention Network (EGAT). The experimental results demonstrate that our proposed method effectively improves the path planning capacity in various application circumstances, outperforming existing methods.