DexPort collects dexterous-hand demos without a full robot setup
Original title: DexPort: A Hand-Held Skill Interface for Dexterous Hand Data Collection
Behavior cloning on more than 2,500 demos raised task success from 36% to 78% when in-the-wild data was included.
DexPort is a portable interface that uses the same high-DoF hand hardware for collection and deployment, recording commanded joint targets and measured joint states. An operator specifies wrist motion and selects skills that coordinate finger motion, so each joint is not controlled independently. The authors collected more than 2,500 demonstrations across laboratory and in-the-wild settings and trained visuomotor policies by behavior cloning without skill labels. On four tasks, using recorded robot commands instead of states raised success from 11% to 64%, and in-the-wild data raised it from 36% to 78%.
- Evaluated on pick-and-place, screwing, drawer stowage, and bimanual scanning.
- Demos can be collected without a robot arm in unstructured environments.
- Policies train on wrist motions and hand commands, not skill labels.
- Commanded targets outperformed measured joint states, 64% versus 11%.
From the original article
1 min read at arxiv.orgCollecting high-quality demonstrations remains a central bottleneck for learning dexterous manipulation. Teleoperation records executable robot actions, but it requires a complete robot setup and places the burden of coordinating many finger joints on the operator.
By Amber Xie · 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.