Learning Push-Grasping in Dense Clutter

Robotic grasping in highly cluttered environments remains a challenging task due to the lack of collision free grasp affordances. In such conditions, non-prehensile actions could help to increase such affordances. We propose a multi-fingered push-grasping policy that creates enough space for the fingers to wrap around an object to perform a stable power grasp, using a single primitive action. Our approach learns a direct mapping from visual observations to actions and is trained in a fully end-to-end manner. To achieve a more efficient learning, we decouple the action space by learning separately the robot hand pose and finger configuration. Experiments in simulation demonstrate that the proposed push-grasping policy achieves higher grasp success rate over baselines and it can generalize to unseen objects. Furthermore, although training is performed in simulation, the learned policy is robustly transferred to a real environment without a significant drop in success rate.