A Deep Reinforcement Learning-Based Motion Planner for Human-Robot Collaboration in Pathological Experiments

摘要

This paper presents a robotic motion planner to address the collision issue between the robotic arm and human arm in human-robot collaboration tasks. We transform the motion planning problem of the robotic arm into a Markov decision process based on real pathological experimental scenarios, by constructing models such as state space, action space, and reward function, to achieve motion planning for a pre-designed task. The planner extracts the position information of the human arm from existing trajectory data sets to generate a real-time hand model, which serves as an obstacle in the collaborative workspace. Based on the reinforcement learning algorithm, the robotic arm adjusts its behavior towards the hand obstacle through a self-designed reward function. Meanwhile, we modify the output action policy of the motion planner by adding OU-Noise to enhance the robotic arm’s exploration capability of the overall space. We deploy the pybullet simulation environment on an Ubuntu20 system and conduct multiple experiments using the CR5 collaborative robotic arm model. Experimental results show that this method can efficiently and accurately plan the motion trajectory of the robotic arm in complex collaboration scenarios, achieving avoidance behavior towards the human arm.