Multi-robot Navigation with Graph Attention Neural Network and Hierarchical Motion Planning

In multi-robot navigation tasks, the interaction between a large number of robots in dynamic environments greatly affects the results of navigation. The interaction between the robots changes with environmental conditions. Therefore, capturing the attention to other robots can greatly improve navigation efficiency. Besides, the learning policy conservatively deals with many high-frequency scenarios. However, in some infrequent scenes, such as dead corners, it can’t perform well. In this paper, we propose a collision avoidance policy trained with deep reinforcement learning, which captures the relationship between robots using Graph Attention Network (GAT). The attention mechanism can indicate the importance of interaction between robots. We present a hierarchical structure to improve the navigation efficiency, which uses motion selector as high-level action and uses collision avoidance policy and target-driven policy as low-level actions. We conduct experiments in Stage simulator and Openai gym, the results indicate that our approach performs better in navigation tasks compared with the state-of-the-art algorithms in the multi-agent field.