DONIC enables one policy for diverse bimanual dexterous manipulation from human data

Dexterous manipulation with two robot hands requires precise coordination of finger motions and contacts to manipulate objects, a capability important for household and industrial tasks. Unlike simpler gripper-based manipulation, it demands tracking complex human hand-object motions and retargeting them to robot kinematics.

Existing methods focus mostly on single-hand tasks or limited bimanual interactions, partly due to the scarcity of datasets with sustained two-hand contact and the complexity of coordinating asynchronous hand roles. A generalist bimanual tracker must learn from diverse human demonstrations, handle changing hand roles, and maintain object motion fidelity, which is challenging because hand-pose accuracy alone does not guarantee successful manipulation.

Training a shared policy across heterogeneous bimanual tasks is complicated by varying contact timing and difficulty. Without careful design, learning signals can be biased by clip difficulty, and enforcing contact on both hands simultaneously can mask progress of individual hands. Thus, new reward and training strategies are needed to enable robust generalization across tasks and objects.