Robot Navigation among External Autonomous Agents through Deep Reinforcement Learning using Graph Attention Network

Abstract

Abstract Finding collision-free and efficient paths in an uncertain dynamic environment is a challenge for robot navigation tasks, especially when there are external autonomous agents that also have decision-making abilities in the same environment. This paper develops a novel method based on DRL with graph attention network (GAT) to solve the problem of robot navigation among external autonomous agents (other agents). Specifically, GAT is adopted to describe the robot and other agents as a specific graph, and extract the spatial structural influence features of other agents on the robot from the graph. Multi-head attention mechanism is utilized to calculate the weights of interactions between the robot and other agents. This GAT uses observations of an arbitrary number of other agents in dynamic environments. Furthermore, the proposed method is combined with optimal reciprocal collision avoidance to improve its safety in new environments. Various simulations demonstrate that our method has good performance and robustness in different environments.