LEAP hand policies reach in-hand poses and cut object drops
Original title: Pushing Dexterity and Safety for In-Hand Object Pose Reaching That Works Remarkably Well
Over 4,800 unseen goals, a residual filter cuts drops 75% while still reaching 91.5% of goals.
Researchers train in-hand object pose-reaching policies for a LEAP hand in simulation without human data, using generated grasp datasets and an adaptive friction curriculum. With system identification, policies transfer zero-shot using motion capture or a wrist-mounted depth camera. A learned residual safety filter with adjustable conservativeness is added on top of the task policy. In 13.3 hours of real-robot evaluation across four objects, the state-based policy reaches 95.2% of goals within 2 cm and 0.4 rad.
- Evaluation covers 4,800 unseen goals and diverse wrist orientations.
- The filter still reaches 91.5% of goals after cutting drops by 75%.
- Policies spin a lightbulb, handle out-of-distribution goals, and tolerate fast arm motion.
- The full training recipe is released as a platform.
From the original article
1 min read at arxiv.orgDexterous manipulation with multi-fingered hands is the next frontier of robotics, yet learned hand skills remain limited to simple grasps or basic repeated motions, such as rotating an object. As robots attempt more demanding in-hand maneuvers, safety, such as not dropping the object, becomes critical, yet it is rarely measured.
By Ian Chuang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.