Robot Crowd Navigation Incorporating Spatio-Temporal Information Based on Deep Reinforcement Learning

Ensuring the safe and efficient navigation of mobile robots in crowds is crucial for their wide application in human society. However, robots face many problems in dense and dynamic crowd scenes, such as complex spatial relationships and difficulty accurately modeling time evolution. Therefore, we propose a novel GG-SAC method to capture spatio-temporal information for robot crowd navigation. Firstly, we utilize the graph attention network to capture complex spatial relationships. Secondly, the gate recurrent unit is introduced to model the temporal evolution of the robot-crowd interaction. Finally, we input the fused spatio-temporal features into Soft Actor-Critic algorithm, a graph-based deep reinforcement learning framework, to train navigation policy. Simulation results show that our method outperforms other methods in terms of success rate and social interaction. Furthermore, our method has good generalization ability for crowd size.