Visual-Based Deep Reinforcement Learning for Robot Grasping with Pushing

Robot grasping is a hot topic in the field of intelligent robot. The synergistic effect of robot grasping and pushing is of great help to manipulation tasks of robots. Someone proposed a model-free grasping method for robot based on deep reinforcement learning combined with a pushing and grasping network, which can determine action choices in pushing and grasping based on visual scene states and then obtain rewards. The network was trained through trial and error. Based on this method, we improve the model by adding feature fusion module and attention module to the deep Q-network, and train and test it in a simulated environment. The experiments show that our model has significantly improved grasping success rate and efficiency. Moreover, in complex unknown environments, the grasping performance of our model is also better than that of the original model.