DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations
One two-hand policy (PPO + BC on its own verified successes) tracks retargeted human hand-object motions; ships BiHOI (1,575 sequences, 229 objects). Reports 86.8% on unseen trajectories, 57.3% on new objects.用一个双手跟踪策略（PPO 结合对自身成功执行的行为克隆）复现重定向后的人手-物体交互动作，并发布 BiHOI 数据集（1575 段双手序列、229 个物体）。论文报告在未见轨迹上成功率 86.8%、新物体上 57.3%。
BibTeX
@misc{sun2026donic,
title = {{DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations}},
author = {Mingze Sun and Ziang Cao and Yuan Fei and Qiankuan Tao and Zihan Wang and Tianqi Liu and Junhao Chen and Haiwen Diao and Fangzhou Hong and Zhaoxi Chen and Ziwei Liu},
year = {2026},
eprint = {2610.29997},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.29997},
}Manipulation & Dexterous Hands操作与灵巧手Robot Learning: RL & Imitation机器人学习：强化学习与模仿学习Datasets & Benchmarks数据集与基准Robot arm机械臂Dexterous hand灵巧手simdexterousbimanualtrackingppohuman-demos