DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations
One checkpoint achieved 57.32% success on held-out objects without finetuning, the authors report.
DONIC learns coordinated two-hand manipulation from human demonstrations, using a new dataset of 1,575 sequences spanning 229 objects. Its training combines reinforcement learning with behavioral cloning, feeding verified successful executions back into the demonstrations. Success reached 86.79% on unseen trajectories involving training objects. The reported gains over leading bimanual tracking baselines were 45.73 percentage points on held-out objects and 77.36 points on training objects.
HF Daily Papers' note
DONIC learns coordinated two-hand manipulation from human demonstrations, using a new dataset of 1,575 sequences spanning 229 objects. Its training combines reinforcement learning with behavioral cloning, feeding verified successful executions back into the demonstrations. Success reached 86.79% on unseen trajectories involving training objects. The reported gains over leading bimanual tracking baselines were 45.73 percentage points on held-out objects and 77.36 points on training objects.
HF Daily Papers' note
score 5