Learning Robot Grasping from a Random Pile with Deep Q-Learning

Grasping from a random pile is a great challenging application for robots. Most deep reinforcement learning-based methods focus on grasping of a single object. This paper proposes a novel structure for robot grasping from a pile with deep Q-learning, where each robot action is determined by the result of its current step and the next n steps. In the learning structure, a convolution neural network is employed to extract the target position, and a full connection network is applied to calculate the Q value of the grasping action. The former network is a pre-trained network and the latter one is a critical network structure. Moreover, we deal with the "reality gap" from the deep Q-learning policy learned in simulated environments to the real-world by large-scale simulation data and small-scale real data.