Author(s)

Guoqiang Luan, Renjun Chen, Junqi Zhang, Yutong Han

Corresponding Author

Guoqiang Luan

Abstract

Due to the fact that traditional industrial flexible robotic arms often use rigid materials with high self weight in the connecting rods and joint components to ensure repetitive accuracy at the end, in order to minimize the displacement of the end position caused by the elastic deformation of the material. This paper investigates the automatic control problem of the end position of flexible robotic arms based on the DRL (Deep reinforcement learning) algorithm. Aiming at the problem that DDPG (Deep Deterministic Policy Gradient) algorithm takes a long time in the training process, a DDPG algorithm based on transfer learning is proposed to plan the end position of flexible manipulator. The motion planning problem of flexible manipulator is decomposed into the superposition of the action output of initial motion planning strategy and the action output of DRL strategy to solve the problem of automatic control of the end position of flexible manipulator. The experimental results show that the proposed residual DRL method is basically stable above 90% at the end of 5000 rounds of training. However, the success rate of DDPG algorithm is rising slowly with the training, and the success rate is close to 60% in 5000 rounds. The experimental results show that DDPG algorithm based on transfer learning can effectively improve the efficiency of DRL agent training.

Keywords

Deep reinforcement learning; Flexible manipulator; End position; Automatic control