Efficient Learning of Goal-Oriented Push-Grasping Synergy in Clutter

We focus on the task of goal-oriented grasping, in which a robot is supposed to grasp pre-assigned goal object clutter and needs some pre-grasp actions such as pushes enable stable grasps. However, this task, gets positive rewards from environment only when successfully grasping object. Besides, joint pushing elongates action sequence, compounding problem reward delay. Thus, sample inefficiency remains main challenge task. In letter, goal-conditioned hierarchical reinforcement learning formulation with high efficiency proposed learn push-grasping policy for specific clutter. our work, improved by two means. First, we use mechanism relabeling enrich replay buffer. Second, policies are respectively regarded generator discriminator trained supervision discriminator, thus densifying rewards. To deal distribution mismatch caused different training settings policies, an alternating stage added turn. A series experiments carried out simulation real world indicate that method can quickly effective outperforms existing methods completion rate success less times motion. Furthermore, validate system also adapt goal-agnostic conditions better performance. Note be transferred without any fine-tuning. Our code available at https://github.com/xukechun/Efficient_goal-oriented_push-grasping_synergy