Vision-Based Robotic Grasping in Cluttered Scenes via Deep Reinforcement Learning

The complicated and unpredictable robotic operating environment frequently results in a poor success rate or failure of robotic grasping. This article proposes a scene dispersion deep reinforcement learning (SDDRL) approach by using a dispersion degree of objects to improve the success rate of robotic grasping, in particular for unknown irregular objects in cluttered scenes. Specifically, an end-to-end dispersion degree reward generation network model is designed to evaluate the degree of dispersion of objects in the robotic workspace. Correspondingly, a grasping strategy generation network is developed by optimizing the dispersion degree of objects, which generates strategies for robotic grasping positions and angles, enabling the robot to grasp objects successfully and make the workspace more scattered, so that the efficiency of grasping can be improved greatly. Extensive experiments performed on the CoppeliaSim simulation environment show that the proposed SDDRL outperforms state-of-the-art methods.