Learning Stable In-Grasp Manipulation in a Non-Dropping Action Space
It addresses the inefficiency and instability in reinforcement learning for dexterous manipulation by integrating classical constraints, benefiting robotics researchers seeking sample-efficient skill acquisition.
This paper introduces a method for learning stable in-grasp manipulation by decomposing dexterous skills into simpler components guided by physics and control theory, achieving efficient and stable skill learning under various conditions.
Traditionally, dexterous manipulation controllers are designed using analytic models constrained by strong assumptions about the hand and the objects being manipulated. Reinforcement learning (RL) has become another common approach in which skills are explored openly in an end-to-end manner but is inefficient because of unnoticeable instability and conflicts in learning objectives. This paper attempts to efficiently explore stable and accurate manipulation skills by decomposing dexterous skills into multiple simpler/analyzable components. Each skill component is subsequently learned with constraints and guidance from classical physics and control theory. Our work shows that for stable grasp, in-grasp reposition/reorientation with different objects, sensor/motor noise, latency, and frictional conditions, skill learning becomes efficient and stable with prior knowledge from theory.