DONIC trains one policy for generalist bimanual dexterous tracking
Original title: DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations
Without finetuning, one checkpoint reaches 86.79 percent success on unseen training-object trajectories and 57.32 percent on held-out objects.
DONIC is presented as a generalist bimanual tracker trained jointly across operation categories and human-motion sources. It is paired with BiHOI, a multi-view dataset of 1,575 sequences spanning 229 objects, with reconstructed interactions retargeted as robot references. Training combines proximal policy optimization and behavioral cloning, recycling verified successes into demonstrations. The authors report large gains over prior bimanual tracking baselines without finetuning.
Unconfirmed: Success rates and baseline margins are the authors' reported results.
- BiHOI has 1,575 multi-view sequences spanning 229 objects.
- Training recycles verified successful executions into new demonstrations.
- Object-centered per-hand rewards supervise each hand by its reference contact role.
- Gains over baselines are 77.36 and 45.73 percentage points.
From the original article
1 min read at huggingface.co
Despite recent progress in dexterous reference tracking, existing models focus primarily on single-hand manipulation and a limited range of interactions, leaving the bimanual dexterous behaviors common in real life unaddressed. The scarcity of contact-rich bimanual manipulation data and the increased complexity of coordinating contacts make developing a generalist bimanual tracker particularly challenging.
By Mingze Sun, Ziang Cao, Yuan Fei · huggingface.co
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.