Humanoid motion planning of robotic arm based on human arm action feature and reinforcement learning

Abstract

The use and application of robotic arms in helping the aged and vulnerable persons are increasing gradually. In order to achieve safer and reliable human-robot interaction and its wider adoption, the requirements for the humanoid motion of robotic arms are becoming more stringent. This paper presents a humanoid motion planning method for a robotic arm based on the physics of human arm and reinforcement learning. Firstly, the humanoid motion rules are extracted by analyzing and learning the action data of human arm, which is collected using the VICON optical motion capture system. Then, according to the acquired features and rules, the corresponding reward functions are proposed and the humanoid motion training of the robotic arm is carried out by using the reinforcement learning based on Deep Deterministic Policy Gradient (DDPG) and Hindsight Experience Replay (HER) algorithm. Finally, the experiments are carried out to verify whether the robotic arm motions planned by the proposed approach are humanoid, and the observed results show its feasibility and effectiveness in planning the humanoid motion of the robotic arm.