Socially Aware Hybrid Robot Navigation via Deep Reinforcement Learning

Abstract

Service robots working in public environments require the capacity to navigate among humans and other obstacles safely and socially compliantly. This paper presents a hybrid navigation approach combining rule-based trajectory generators into deep reinforcement learning for motion planning in populated and cluttered environments. An intention-based action space is proposed in the reinforcement learning framework to achieve a tight coupling of the two methods. By fusing static information and dynamic objects, our network can learn motion patterns adapted to real-world scenarios. The rule-based trajectory generator guarantees the safety and dynamic feasibility of the motion primitives. The robot is trained to understand real-time human-robot interactions through deep reinforcement learning. Experiment results demonstrate that our policy can efficiently perceive human interactions and navigate the robot safely in crowded environments with static obstacles.