Reinforcement Learning-based Improved DDPG for Robotic Arm Grasping

Traditional robotic arm gripping control methods, while capable of providing high control accuracy, often rely on expensive precision hardware as well as lack sufficient stability. In order to realize higher accuracy and flexibility of control on a comparatively low-cost robotic arm, this paper proposes an improved DDPG reinforcement learning algorithm to optimise robotic arm grasping control. Firstly, a simplified system of a four-degree-of-freedom robotic arm with a grabber is established as a model framwork in the ROS. Then, for the insufficiency of the reward function of the traditional DDPG algorithm, a composite reward function is designed to improve the learning efficiency. Meanwhile, in order to solve the sampling productivity issues during the gripping process with robotic arm, a large amount of it was added to the experience pool to improve the study rate. The experimental findings indicate that the modified DDPG accelerates the number of successful grasps with the same number of training sets, and the final success rate reaches 68% after the same number of training sets, while the success rate of the unmodified DDPG algorithm is only 26%.