Robot Navigation with Interaction-based Deep Reinforcement Learning

For the scene of dense crowd flow in limited space, it is very important and challenging for the robot to walk through the dense crowd without collision and move to the destination efficiently. As deep reinforcement learning has achieved certain results in human-aware navigation policies, it provides a feasible solution for the robot navigation in dense crowd. But current environment representation method is difficult to represent the intention of human movement, which causes that the policy network cannot make forward-looking decisions. And the previous learning model could not effectively represent any number of pedestrians and maintain stable navigation capability in unfamiliar environment. In this study, we propose a novel model of robot navigation, that is called robot human interaction reinforcement learning (RHIRL). A new environment representation method is proposed which implicitly includes the potential interaction and effectively improves the navigation ability in unfamiliar and dynamic interactive environment. The experiment results show that the proposed model has obvious advantages and excellent navigation performance in dynamic and unfamiliar environment.