Path-Following Navigation in Crowds With Deep Reinforcement Learning

The local navigation with collision avoidance is becoming increasingly important for the mobile robot in the crowd scenario. Previous work mainly concerns its point-to-point local navigation via deep reinforcement learning (DRL). However, applying DRL to the local path-following navigation poses extra challenges in generating smooth trajectory and enhancing safety. This paper presents a danger-aware robot navigation algorithm by defining the pedestrians' danger and introducing a virtual robot about the reference path. The main novelty of this algorithm is that the virtual robot is leveraged to derive the extra action and more sampling waypoints in pursuit of the robot motion smoothness and foresight. Moreover, a priority mechanism is established and incorporated into DRL navigation, so as to enhance safety of robot navigation. Experiments on the path-following social navigation demonstrate that our presented algorithm outperforms the state-of-the-art method in terms of the motion smoothness and the safety via evaluation metrics.