Grasp Trajectory Planning for Vehicle-Mounted Robotic Arm based on Improved DDPG Algorithm

Abstract

In response to the problems of slow convergence and poor control effect when the traditional Deep Deterministic Policy Gradient (DDPG) algorithm is used to train the vehicle-mounted robotic arm, an improved DDPG algorithm is proposed in this paper. According to the control characteristics of the vehicle-mounted robotic arm, the state space, action space, and reward and punishment function are designed. Then the operation area is partitioned, and the baseline pose of each area is taught. A guidance function is designed and introduced to guide the algorithm training with the baseline pose to accelerate the algorithm's convergence. Finally, the traditional reward and punishment function is improved according to the guidance training characteristics. The simulation experiments show that the improved DDPG algorithm has significantly improved the convergence efficiency, planning success rate, and control efficiency compared with the traditional DDPG algorithm, especially since the planning success rate can reach about 93.8%.