Robot Navigation in Crowd Based on Dual Social Attention Deep Reinforcement Learning

Finding a feasible, collision-free path in line with social activities is an important and challenging task for robots working in dense crowds. In recent years, many studies have used deep reinforcement learning techniques to solve this problem. In particular, it is necessary to find an efficient path in a short time which often requires predicting the interaction with neighboring agents. However, as the crowd grows and the scene becomes more and more complex, researchers usually simplify the problem to a one-way human-robot interaction problem. But, in fact, we have to consider not only the interaction between humans and robots but also the influence of human-human interactions on the movement trajectory of the robot. Therefore, this article proposes a method based on deep reinforcement learning to enable the robot to avoid obstacles in the crowd and navigate smoothly from the starting point to the target point. We use a dual social attention mechanism to jointly model human-robot and human-human interaction. All sorts of experiments demonstrate that our model can make robots navigate in dense crowds more efficiently compared with other algorithms.