Vision-based Reinforcement learning: Moving object grasping with a Single Active-view Camera

Abstract

Traditional methods of grasping moving objects use visual information from camera that is statically mounted overlooking the task environment. However, when the camera view is hindered by an obstacle or a robot itself performing tasks, these methods cannot produce good results for grasping moving objects. This paper proposes a new moving object grasping method using only single active RGB camera attached to the robot with utilization of model free deep reinforcement learning strategy with Soft Actor-Critic(SAC) algorithm. Our proposed system uses RGB image data to train our proposed system and to further improve training and success rate, we integrate vision data with robot’s kinematic state information during training. A dense reward function is designed for efficient learning of the agent and training is done on simulation using a Ufactory xArm 7 manipulator. Experimental results demonstrate that with our proposed method, the manipulator can learn to autonomously grasp the moving object.