Learning Stable In-Grasp Manipulation in a Non-Dropping Action Space
2026-06-26 • Robotics
Robotics
AI summaryⓘ
The authors explain that controlling a robot hand to manipulate objects is usually done either by relying on strict physical models or by letting the robot learn through trial and error, which can be slow and unstable. They propose breaking down complex hand skills into smaller, simpler parts and teaching each part using rules from physics and control theory. This approach helps the robot learn to grasp and move objects more steadily and efficiently, even when there are real-world challenges like noise and friction.
dexterous manipulationreinforcement learningcontrol theorygrasp stabilitysensor noisemotor noisephysics-based modelingrobot handskill decompositionlatency
Authors
Ha Thang Long Doan, Hikaru Arita, Kazuto Nakashima, Kenji Tahara
Abstract
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.