Cooperative Path Planning of Multi-Agent Based on Graph Neural Network
Hanqi Dai, Weining Lu, Jun Yang, Bin Liang · 2022 34th Chinese Control and Decision Conference (CCDC) · 2022
The objective of the Cooperative Multi-agent Planning problem is to provide each agent with an accessible, efficient and collision-free path, and the key is effective communication and cooperation between agents. Scope of this paper is to present a novel cooperative approach for multi-agent coordination in planning. We presented a combined architecture for Multi-Agent path planning in an unknown environment, and each agent only has local communication and local observations to collaborate and share perceived information with others. The architecture consists of a convolutional neural network (CNN) that extracts adequate features from local perception, a Graph Sample and AggreGate (GraphSAGE) that fuses the features among agents, and a Multilayer Perceptron (MLP) that decodes the output of GraphSAGE into action primitives. We trained the combined neural networks and tested them, the overall efficacy of the presented method was comprehensively evaluated by multiple simulation results in an unknown environment.