Simulation of Real-time Collision-Free Path Planning Method with Deep Policy Network in Human-Robot Interaction Scenario

Collision avoidance is essential for manipulators working in human-robot interaction scenarios where humans and robots collaborate to ensure human safety and working efficiency. However, traditional path planning methods like RRT are not well-suited for the dynamic and real-time requirement because of the human intervention where the human arms move continuously and randomly. Therefore, we propose a deep policy network based on reinforcement learning to realize real-time collision-free trajectory planning. First, we use a state-of-art human pose estimation network for robot perception. Second, a reinforcement learning-based path plan method is designed to realize the real-time obstacle avoidance. Also, We create the simulation environment based on the human-robot interaction scenarios and the training results show that our method can realize real-time obstacle avoidance path planning in human-robot interaction scenarios by offline training and online deploying. To see the simulation results, please visit the project video at https://www.bilibili.com/video/BVlbW4ylq70Q.