Pushing Dexterity and Safety for In-Hand Object Pose Reaching That Works Remarkably Well
Pushing Dexterity and Safety for In-Hand Object Pose Reaching That Works Remarkably Well
This research introduces a method for training dexterous manipulation policies for a LEAP hand in simulation, without human data. The approach combines diverse resets, generated grasp datasets, and an adaptive friction curriculum. Policies transfer zero-shot to the real hand using motion capture or a wrist-mounted depth camera, achieving 95.2% success in reaching goals within 2 cm and 0.4 rad. A learned safety filter reduces object drops by 75% while maintaining 91.5% success. The work also demonstrates the ability to spin a lightbulb, reach out-of-distribution goals, and withstand perturbations, with the full recipe released as a platform for studying high-dimensional dexterous systems.
This paper presents a novel approach to training dexterous manipulation policies for a LEAP hand, emphasizing both performance and safety. The method combines a diverse grasp dataset, adaptive friction curriculum, and a residual safety filter, achieving high success rates in real-world tasks with minimal hardware.
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
real-robot evaluation hours
13.3 hours
View original evidence
culum. With careful system identification, the policies transfer zero-shot to the real hand, using either motion capture or a single wrist-mounted depth camera, and prove remarkably capable and robust. On top of the task policy, we add a learned residual safety filter with adjustable conservativeness, letting us tune the balance between performance and safety and greatly reduce object drops. In 13.3 hours of real-rob
Open source S4unseen goals
>4,800
View original evidence
ot 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, and the filter cuts its 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 p
Open source S5objects used in evaluation
4
View original evidence
ot 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, and the filter cuts its 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 p
Open source S5wrist orientations tested
4
View original evidence
ot 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, and the filter cuts its 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 p
Open source S5tasks performed by robot
2
View original evidence
ot 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, and the filter cuts its 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 p
Open source S5
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
culum. With careful system identification, the policies transfer zero-shot to the real hand, using either motion capture or a single wrist-mounted depth camera, and prove remarkably capable and robust. On top of the task policy, we add a learned residual safety filter with adjustable conservativeness, letting us tune the balance between performance and safety and greatly reduce object drops. In 13.3 hours of real-rob
Open source S4
ot 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, and the filter cuts its 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 p
Open source S5
ortable skills such as screwing in a lightbulb. We release the full recipe 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/.
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org