Abstract:
Grasping a target object in the cluttered environment is challenging due to potential collisions. Taking pre-grasp manipulations such as pushing, sliding and poking is an...Show MoreMetadata
Abstract:
Grasping a target object in the cluttered environment is challenging due to potential collisions. Taking pre-grasp manipulations such as pushing, sliding and poking is an effective way to singulate the target. However, the success rate is heavily affected by the dimension disaster derived from complex scenarios. To address the problem, we propose a multi-teachers distillation strategy which includes two stages. At the first stage, we manually decompose the complex singulation goal into strongly and weakly related subtasks that are easy to reach, and let one teacher completes lesson preparation of one subtask. At the second stage, one student who can finish the whole task is distilled by directly learning from teachers-experienced. Moreover, for the lesson preparation of strongly related subtask, deep transfer learning is leveraged to accelerate teachers' training on the low-level teacher knowledge, and invalid action suppression strategy is effectively incorporated into teachers' learning, by limiting the repetition of empty pushes, to increase sample efficiency. Experiments demonstrate that the proposed method outperforms state of the art methods on singulation success rate, as well as the number of policies especially. Additionally, our method demonstrates superior scalability, and can effectively address unseen scenarios.
Published in: IEEE Robotics and Automation Letters ( Volume: 9, Issue: 11, November 2024)