Dexterous Robot Manipulation from Human Demonstrations via Contact-Anchored Retargeting and Residual Policy Learning
Dexterous Robot Manipulation from Human Demonstrations via Contact-Anchored Retargeting and Residual Policy Learning
This research paper introduces a method to learn dexterous robot manipulation from human demonstrations by focusing on contact structures rather than joint motions. The proposed pipeline uses physics refinement with a simulated MANO hand, contact-anchored retargeting, and residual policy learning to convert human motion-capture recordings into physically consistent trajectories adaptable to different robot hands. Results show substantial improvements in task success rates for both single-hand and dual-hand tasks, with the ability to transfer one human dataset to four morphologically distinct robot hands and execute bimanual tasks on physical hardware without real-robot training data.
The study presents a three-stage pipeline that converts human motion-capture data into dexterous robot manipulation policies without requiring real-robot training data, achieving significant success rate improvements on both single-hand and dual-hand tasks through contact-anchored retargeting and residual policy learning.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
dual-hand tasks
25
Reported trials
View original evidence
ion. The pipeline reconstructs 25,454 single-hand trajectories (success 7.3% -> 59.3%) and 25 dual-hand tasks (16.0% -> 62.4%) with one shared policy per setting, transfers one human dataset to four morphologically distinct robot hands (+62.4 pp), and executes four contact-rich bimanual tasks on physical hardware with zero real-robot training data.
Open source S5bimanual tasks executed
4
Reported trials
View original evidence
ion. The pipeline reconstructs 25,454 single-hand trajectories (success 7.3% -> 59.3%) and 25 dual-hand tasks (16.0% -> 62.4%) with one shared policy per setting, transfers one human dataset to four morphologically distinct robot hands (+62.4 pp), and executes four contact-rich bimanual tasks on physical hardware with zero real-robot training data.
Open source S5
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ion. The pipeline reconstructs 25,454 single-hand trajectories (success 7.3% -> 59.3%) and 25 dual-hand tasks (16.0% -> 62.4%) with one shared policy per setting, transfers one human dataset to four morphologically distinct robot hands (+62.4 pp), and executes four contact-rich bimanual tasks on physical hardware with zero real-robot training data.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org