DONIC teaches one policy to control two hands: 86.79% success without fine-tuning

A team of researchers has released DONIC, a single policy for controlling two dexterous robotic hands that tracks human motion while handling objects. With no fine-tuning, one checkpoint succeeds on 86.79% of unseen trajectories with familiar objects and on 57.32% with objects that were not in the training data. That is a lead of 77.36 and 45.73 percentage points, respectively, over the best existing bimanual tracking baselines.

Subtract those gaps from the results and the competitors land at roughly 9% on familiar objects and about 12% on new ones. In other words, until now bimanual tracking has worked at a level close to "almost never."

Why two hands are harder than one hand times two

Until now, reference-motion tracking for dexterous hands has mostly centered on a single hand and a narrow set of interactions. Real life is different: unscrewing a lid, holding a box from below and from the side, passing an object from one palm to the other. All of this requires the two hands to play different roles and swap them at the right moment.

There are two problems here. First, data on contact-rich bimanual manipulation is very scarce. Second, each additional hand sharply increases the difficulty of coordinating contacts: a mistake by one hand breaks the other's grasp. That is why a general-purpose bimanual tracker has not existed until now.

BiHOI: the data that was missing

The authors started with the dataset. BiHOI is a multi-view collection of 1575 sequences with 229 objects, focused on sustained two-handed interaction rather than brief touches. Interactions reconstructed from video are retargeted into robot references, meaning human motions are converted into trajectories that a robotic hand can execute. DONIC was trained simultaneously on different manipulation categories and different sources of human motion, and this mix matters for generalization.

A loop in which the robot learns from its own successes

Training combines proximal policy optimization with behavioral cloning. The two work in a feedback loop: rollouts that pass a success check are fed back into training as demonstrations. The policy itself replenishes the pool of good examples.

Two other design decisions concern rewards and advantage estimation:

• Object-centered per-hand reward. The reward is tied to the object's motion, so feedback on what happens to the object is preserved. At the same time, each hand is evaluated according to its contact role in the reference. • Clip-Centered Advantages. The method removes advantage biases that depend on the specific reference but keeps the differences within a trajectory. The reasoning: clips vary considerably in difficulty, and without centering, training would drift toward the easy ones.

Together, these decisions let a single policy learn coordinated manipulation across heterogeneous references.

Key parameters

• The first general-purpose bimanual tracker, trained at once on several manipulation categories and motion sources • BiHOI: 1575 sequences, 229 objects, multi-view • 86.79% success on new trajectories with familiar objects • 57.32% on held-out objects • Lead over the best baselines: 77.36 and 45.73 percentage points • One checkpoint, no fine-tuning

Who will find this useful first

People building humanoids and manipulators with dexterous hands. The bottleneck is not the hardware but data and control: if human demonstrations can be reliably translated into bimanual references, and a single policy executes them, then collecting data through video and teleoperation becomes a working pipeline. Such a policy can serve as a low-level executor, with a planner or a language model on top that specifies what exactly to do.

The limitation is also visible in the numbers. Success on new objects is 57.32%, which means the system still fails in nearly every second case. That is not enough for a factory or a home, but as a research foundation and a benchmark for generalization, the result is strong. Generalization to the shape and weight of unfamiliar objects can now be measured and improved on a shared dataset with a clear baseline.