Socially Compliant Navigation in Indoor Corridors Based on Reinforcement Learning

In this paper, the policy function of a mobile robot navigating in indoor corridor environments was obtained through reinforcement learning (RL). Assuming the scenario where right-passing rules are enforced, target paths associated with different corridor widths were defined; the robot's actions of navigating into the target paths were defined as RL rewards to encourage the common social consensus regarding corridor passing. Specifically, the robot was trained to render proper reactions according to the width of the corridor and the robot's speed, pose, and relative position in the corridor. The RL models were trained in Gazebo and ROS; the effectiveness of the navigation policy was validated by various tests of different conditions. It was found that different speeds need different strategies; the RL models trained for each specific speed category appear to be optimal. Such results were supported by the cross-examinations on the success rate and the number of corrective actions.