A Method of Robot Grasping Based on Reinforcement Learning

Abstract

In this paper, we propose a method for robot grasping based on reinforcement learning. A six-degree-of-freedom robot is used to execute the grasping action, and an RGB-D camera is used as a sensing module to collect the information. Different from previous model-based methods, our whole system is data-driven. In order to better grasp a bunch of randomly placed objects, the pushing action is added to spare more space for grasping. The whole process is described as a Markov decision process (MDP). The training is carried out in simulation. Experimental results show that our method can complete the grasping work efficiently, especially in those complicated and difficult scenarios. At the same time, the model can be directly applied in actual grasping tasks which greatly reduce the risk of training in the physical world.