Ropedia's DONIC lifts bimanual tracking success to 86.79%

A single reinforcement-learning policy, trained jointly across human-demonstration sources with no per-task fine-tuning, is reported to track bimanual dexterous manipulation on unseen trajectories at 86.79% success, up from 9.43% for the strongest prior bimanual baseline. The system is called DONIC and comes from a team at Ropedia with corresponding author Ziwei Liu; the paper is posted on Hugging Face.

The abstract claims the model achieves "86.79% success on unseen trajectories of training objects and 57.32% on held-out objects, surpassing the state-of-the-art bimanual tracking baselines by 77.36 and 45.73 percentage points." Wrist translation error on unseen objects falls from 19.46 cm for Dexplore-bi to 4.58 cm.

The second half of the release is BiHOI, a multi-view dataset of 1,575 sequences across 229 objects, captured with 32 calibrated RGB cameras at 1920×1232 and 30 fps over 14.77 hours of interaction. Both hands are in simultaneous contact with the object on 65.91% of contact-active frames; the next-highest number the authors tabulate for a compared public dataset is 38.57% for H2O.

Training runs in Isaac Gym with 8,192 parallel environments, combining proximal policy optimization with behavioral cloning in a feedback loop that recycles verified successful executions back into the demonstration pool. The reported figures are from simulation; the abstract does not describe a real-robot deployment. The release lands into a busy run of generalist-controller work in robotics, with 168 stories tracked in our feed over the last 90 days.