AdaDex turns short human demos into adaptable dexterous skills
Original title: Learning Adaptable Dexterous Skills via Hybrid Dexterous Mimicking
One policy trains on motion tracking and noised goal environments, then transfers to real robots.
Researchers present AdaDex, a goal-conditioned reinforcement-learning framework that converts a few seconds of hand-object motion into an adaptable multi-finger skill. Hybrid Dexterous Mimicking trains a shared policy on motion-tracking environments and on generalization environments with noised states, goals, and a sparse object-success reward. Reported skills include functional grasping, reposing, in-hand reorientation, and throw-and-catch. The authors say it outperforms baselines and transfers across simulators and to the real world.
- Input can come from motion capture or reconstructed in-the-wild RGB(D) video.
- Tracking environments use a dense motion-tracking objective.
- Generalization environments use a sparse object-success reward.
- The authors report simulator-to-simulator and real-world transfer.
From the original article
1 min read at arxiv.orgDexterous multi-finger hands hold promise to unlock diverse skills like functional grasps, in-hand reorientation, and dynamic handover that the human-designed world assumes, yet existing learning recipes force a trade-off.
By Yuxuan Kuang · 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.