Robotic Grasping based on Point Cloud Recognition and Deep Reinforcement Learning

We propose a deep reinforcement learning method to learn to produce grasping strategies using 3D point clouds as input. We facilitate a training process by using a tensor-based distributed training framework to perform many trial-and-errors with a wide variety of objects. Our approach greatly increases the number of object categories that can be handled and exhibits a strong generalization ability to grasp unknown object categories.