Pushing Dexterity and Safety for In-Hand Object Pose Reaching That Won't Drop the Object

Forum topic · 小凯 · 2026-10-28

Summary

This paper from UC Berkeley-affiliated researchers (Ian Chuang, Hiya Shah, Vamsi Chilamkur, Iman Soltani, Claire Tomlin, arXiv:2610.37652) addresses in-hand object pose reaching on a LEAP robotic hand: moving an object to nearly arbitrary positions and orientations through long sequences of consecutive goals. The team trains policies entirely in simulation without human data, using generated grasp datasets for diverse resets and goal sampling, an adaptive friction curriculum, and careful system identification. Policies transfer zero-shot to the real hand using motion capture or a single wrist-mounted depth camera. A learned residual safety filter with adjustable conservativeness trades performance against safety, reducing object drops. Over 13.3 hours of real-robot evaluation across 4,800 unseen goals, 4 objects, and diverse wrist orientations, the state-based policy reached 95.2% of goals within 2 cm and 0.4 rad tolerance; the filter cut drops by 75% while still reaching 91.5% of goals. Demonstrated skills include spinning a lightbulb in arbitrary poses, reaching out-of-distribution goals, and withstanding fast arm motion and perturbations. The full training recipe is released as a platform for studying performance, reliability, and safety in high-dimensional dexterous systems.

Paper Overview

Field: Robotics / Dexterous Manipulation

Authors: Ian Chuang, Hiya Shah, Vamsi Chilamkur, Iman Soltani, Claire Tomlin

arXiv: 2610.37652

Project page: https://ian-chuang.github.io/dexsafe/

Abstract

Dexterous 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.

The authors study in-hand object pose reaching, a task that demands diverse and difficult grasps: moving an object to almost any position and orientation within reach of the fingers, at any wrist orientation, through long sequences of consecutive goals.

Key Contributions

A recipe for training policies for a LEAP hand in simulation without human data, built on insights including:
Diverse resets and goals via generated grasp datasets
An adaptive friction curriculum
Careful system identification
Zero-shot transfer to the real hand using either motion capture or a single wrist-mounted depth camera
A learned residual safety filter with adjustable conservativeness on top of the task policy, allowing tuning of the balance between performance and safety

Results

In 13.3 hours of real-robot evaluation over 4,800 unseen goals across 4 objects and diverse wrist orientations:

The state-based policy reaches 95.2% of goals within 2 cm and 0.4 rad
The safety filter cuts drops by 75% while still reaching 91.5% of goals

The policies can also spin a lightbulb held in an arbitrary pose, reach out-of-distribution goals, withstand fast arm motion and perturbations, and—needing only a wrist camera—serve as portable skills such as screwing in a lightbulb.

The full recipe is released as a platform for studying performance, reliability, and safety in high-dimensional dexterous systems. Videos and more are available at https://ian-chuang.github.io/dexsafe/.