Deep Reinforcement Learning-based Trajectory Planning for Manipulator Obstacle Avoidance

Abstract

A deep reinforcement learning (DRL)-based motion planning method is proposed to improve long planning elapse and lengthy path of the traditional planning algorithms for robotic manipulator movement in obstacle avoidance. Firstly, based on the mathematical model of the manipulator and the motion environment, the DOBOT robot and the operating environment are built in PyBullet, and the parameters such as the reward function, the action and the state variables required for DRL are set. Secondly, the deep deterministic policy gradient (DDPG) algorithm is applied for the characteristics of static obstacle avoidance, and motion simulation experiments are conducted. The simulation results show that the proposed DDPG algorithm has a certain degree of improvements in planning elapse and path length compared with the rapid-exploring random tree (RRT) algorithm and the improved RRT algorithm. Finally, the effectiveness of the DDPG algorithm in obstacle avoidance operations is tested using the DOBOT robot in a laboratory environment with multiple obstacles.

Keywords

• Manipulator;Deep reinforcement learning;Obstacle avoidance path planning;Deep deterministic policy gradient algorithm