Reproduce “DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mingze Sun, Ziang Cao, Yuan Fei, Qiankuan Tao, Zihan Wang, Tianqi Liu, Junhao Chen, Haiwen Diao and 3 more
DONIC trains a shared policy for bimanual reference-conditioned object manipulation, using object-motion feedback, reference-defined contact roles, and clip-wise advantage centering. BiHOI contains 1,575 sequences across 229 objects. In simulation, one checkpoint without fine-tuning achieves 57.32% success on 164 held-out-object clips, versus 11.59% for an adapted baseline; this establishes tracking only for tested references, not physical deployment or arbitrary goals.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.