Moving Object Grasping Method of Mechanical Arm Based on Deep Deterministic Policy Gradient and Hindsight Experience Replay

Abstract

The mechanical arm is an important component in many types of robots; however, in certain production lines, the conventional grasp strategy cannot satisfy the demands of modern production because of several interference factors such as vibration, noise, and light pollution. This paper proposes a new grasping method for manipulators in stamping automatic production lines. Considering the factors that affect grasping in the production environment, the deep deterministic policy gradient (DDPG) method is selected in this study as the basic reinforcement-learning algorithm, and this algorithm is used to grasp moving objects in stamping automatic production lines. Owing to the low success rate of the conventional DDPG algorithm, the hindsight experience replay (HER) is used to improve the sample utilization efficiency of the agent and learn more effective tracking strategies. Simulation results show an 82% mean success rate of the optimized DDPG-HER algorithm, which is 31% better than that of the conventional DDPG algorithm. This method provides ideas for the research and design of the sorting system used in stamping automation production lines.