Moving Object Flexible Grasping Based on Deep Reinforcement Learning

Abstract

High manipulability can improve the movement ability of the robot end-effector in various degrees of freedom when it is engaged in autonomous grasping. In this paper, we propose a flexible grasping method based on deep reinforcement learning for moving objects. We use the deep deterministic policy gradient (DDPG) algorithm to train and control a six degrees of freedom manipulator, and introduce the manipulability index to optimize the grasping pose of the manipulator. A moving ball grasping mission was conducted using the Robot Operating System (ROS) and the Gazebo simulator to verify the effectiveness of the method. Compared with the DDPG algorithm without optimizing manipulability and the traditional tracking method, experimental results indicate that the proposed method maintains the high manipulability of the manipulator.