Robot hand learns nimble moves with fewer dropped objects
Robot hand learns nimble moves with fewer dropped objects
Researchers at UC Berkeley and Davis trained a 16-joint LEAP robotic hand in simulation without human demonstrations, then added a separate safety filter to reduce object drops from 16 to 4 over 1,200 movement sequences. The approach allows adjusting caution without retraining the underlying skill, though it slightly reduces target-reaching success rates. Tests used external motion-tracking cameras and included real objects like lightbulbs and 3D-printed items.
The study presents a simulation-trained robot hand controller augmented by a separate safety filter that reduces object drops from 16 to 4 over 1,200 sequences but lowers target-reaching success from 95.2% to 91.5%, demonstrating an adjustable trade-off between dexterity and grip retention.
Source: Tech Xplore — Robotics · Read original article ↗
Article text · Original source · English
Manipulating an object with one hand is a balancing act. Fingers must release and renew their grip while keeping the object from slipping away, even with the palm facing downward.
Researchers have now trained a robot hand to perform these maneuvers and added a separate software safeguard to reduce object drops. In tests guided by external motion-tracking cameras, the safeguard cut drops from 16 to four over 1,200 requested movement sequences. The trade-off was that the hand missed a few more targeted motions.
Ian Chuang and colleagues at the University of California, Berkeley and Davis describe their work in a paper posted to the arXiv preprint server. Their approach lets researchers adjust how cautiously the hand moves without retraining its underlying skill.
Practice first with an easier grip
The team trained a 16-joint LEAP robotic hand entirely in simulation, without human demonstrations. It had to shift objects to different positions and orientations between its fingers, going beyond just repeatedly spinning them.
Training began with hundreds of thousands of computer-generated grasps. The simulated fingertips initially had unrealistically high friction, making the objects easier to hold. As the model's performance improved, that assistance was gradually decreased to approximate more realistic conditions. The researchers also measured a real hand's motor behavior and incorporated it into the simulation.
Tests were run using a real lightbulb and plastic, 3D-printed versions of a knife, wineglass and duck. Each object required its own controller. The safeguard was trained separately to adjust the main controller's proposed movements. Its job was to prevent drops while changing those movements as little as possible.
The safety filter did not see the goal position or earn rewards for reaching it. "This separates learning the task from learning safety," the authors write.
Keeping hold can mean missing the target
The evaluation included 4,800 requested movements of the four objects using six wrist orientations. Tests came in sequences of 50 targets, and the hand was reset if the object fell or became stuck.
With external cameras tracking the object, the robotic hand reached 95.2% of targets without the safeguard and 91.5% with it. A task was counted as successful if the object came within 2 centimeters and about 23 degrees of its target within 10 seconds.
Although the safeguard reduced overall target-reaching success, it allowed the hand to complete more movements before its first drop. That distinction matters when a dropped object interrupts a task.
The team also demonstrated screwing a lightbulb into a socket until it lit, combining learned finger movements with scripted arm motions. The safeguard did not improve success in those small demonstrations, showing that its benefits did not automatically extend to every task.
For now, the system offers an adjustable compromise between reaching a manipulation target and keeping a grip on an object. Making that protection work reliably and training a single controller to manipulate multiple objects present the next challenge.
Written for you by our author Andrew Zinin, edited by Lisa Lock, and fact-checked and reviewed by Robert Egan—this article is the result of careful human work. We rely on readers like you to keep independent science journalism alive. If this reporting matters to you, please consider a donation (especially monthly). You'll get an ad-free account as a thank-you.
Publication details
Ian Chuang et al., Pushing Dexterity and Safety for In-Hand Object Pose Reaching That Works Remarkably Well, arXiv (2026). DOI: 10.48550/arxiv.2610.37652
Project site: ian-chuang.github.io/dexsafe/
Journal information: arXiv
Who's behind this story?
Andrew Zinin
Master's in physics with research experience. Long-time science news enthusiast. Plays key role in Science X's editorial success. Full profile →
Lisa Lock
BA art history, MA material culture. Former museum editor, paramedic, and transplant coordinator. Editing for Science X since 2021. Full profile →
Robert Egan
Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile →
© 2026 Science X Network
Citation: Robot hand learns nimble moves with fewer dropped objects (2026, November 3) retrieved 3 November 2026 from https://techxplore.com/news/2026-11-robot-nimble.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S4
Reported numbers
requested movement sequences
>1,200
Reported trials
View original evidence
educe object drops. In tests guided by external motion-tracking cameras, the safeguard cut drops from 16 to four over 1,200 requested movement sequences. The trade-off was that the hand missed a few more targeted motions. Ian Chuang and colleagues at the University of California, Berkeley and Davis describe their work in a paper posted to the arXiv preprint server. Their approach lets researchers adjust how cautiousl
Open source S3evaluation movements
4,800
Reported trials
View original evidence
ntation. One policy per object; the outline is the goal and turns green when reached. Sped up only once every object holds its goal. Keeping hold can mean missing the target The evaluation included 4,800 requested movements of the four objects using six wrist orientations. Tests came in sequences of 50 targets, and the hand was reset if the object fell or became stuck. With external cameras tracking the object, the r
Open source S7
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
y the hand moves without retraining its underlying skill. Practice first with an easier grip The team trained a 16-joint LEAP robotic hand entirely in simulation, without human demonstrations. It had to shift objects to different positions and orientations between its fingers, going beyond just repeatedly spinning them. Training began with hundreds of thousands of computer-generated grasps. The simulated fingertips i
Open source S4
obotic hand reached 95.2% of targets without the safeguard and 91.5% with it. A task was counted as successful if the object came within 2 centimeters and about 23 degrees of its target within 10 seconds. Although the safeguard reduced overall target-reaching success, it allowed the hand to complete more movements before its first drop. That distinction matters when a dropped object interrupts a task. The team also d
Open source S8
educe object drops. In tests guided by external motion-tracking cameras, the safeguard cut drops from 16 to four over 1,200 requested movement sequences. The trade-off was that the hand missed a few more targeted motions. Ian Chuang and colleagues at the University of California, Berkeley and Davis describe their work in a paper posted to the arXiv preprint server. Their approach lets researchers adjust how cautiousl
Open source S3
ntation. One policy per object; the outline is the goal and turns green when reached. Sped up only once every object holds its goal. Keeping hold can mean missing the target The evaluation included 4,800 requested movements of the four objects using six wrist orientations. Tests came in sequences of 50 targets, and the hand was reset if the object fell or became stuck. With external cameras tracking the object, the r
Open source S7
Source：Tech Xplore — Robotics · techxplore.com