DONIC Explained: Teaching Two Robot Hands From Human Demonstrations

Ropedia published DONIC on October 21, 2026, a research system that teaches one policy to control two dexterous robot hands from human demonstrations. It improves tracking on unfamiliar movements and objects in simulation. The release also introduces BiHOI, a dataset designed around sustained two-handed contact.

The useful change is coordination. When one hand holds an object and the other moves it, treating each hand as an independent problem can miss what makes the action work.

How DONIC learns two-handed manipulation

DONIC follows reconstructed motion references. It does not take an arbitrary instruction and independently decide how to perform a household task.

The project's method explanation describes a shared controller for both hands. An object-centered reward tracks the object's movement while evaluating each hand according to the contact role required by the reference. This gives the system a way to distinguish a supporting hand from a manipulating hand.

Training combines reinforcement learning with behavior cloning. Successful policy executions are replayed to verify them, then reused as demonstrations in later training. Another component, Clip-Centered Advantages, adjusts the learning signal across reference clips so their different baselines do not dominate updates.

The broader idea is practical: a demonstration library can improve as the controller learns, rather than staying fixed at the quality of the initial reconstructed motions.

DONIC benchmark results and their limits

The technical report evaluates a single checkpoint without fine-tuning on the held-out tests. Its main results separate unfamiliar trajectories of known objects from unfamiliar objects.

Training data	Unseen trajectories	Unseen objects
Public human-motion datasets	86.79% success	57.32% success
Public datasets plus BiHOI	88.68% success	59.76% success

These are the authors' simulation results. The comparison uses bimanual adaptations of DexTrack and Dexplore, not a universal leaderboard of commercial robot systems. Both hands must satisfy the tracking thresholds for a clip to count as successful.

The demos are Isaac Gym policy rollouts, not physical-robot demonstrations. The paper limits its quantitative claims to simulated single-object interactions. Transfer to physical hardware, longer tasks, multiple objects and an independent benchmark remain future work.

That boundary matters more than the headline success rate. Contact behavior depends on details such as the object representation and the motion reference. A strong simulator result does not establish the same reliability with real friction, sensing errors and unexpected objects.

What BiHOI adds

BiHOI contains 1,575 interaction sequences across 229 objects and 14.77 hours of activity. The capture setup uses 32 calibrated cameras to reconstruct the hands and object before translating human movements into robot references.

The emphasis is sustained two-hand interaction. That makes it relevant to the part of manipulation that a large collection of isolated grasps might still miss: maintaining one contact while the other hand changes the object's state.

The improved held-out results suggest that more suitable demonstrations help the controller generalize. They do not show that adding arbitrary video would have the same effect. Here, the footage goes through reconstruction and retargeting before it becomes a training reference.

Code, checkpoints and dataset availability

The official repository contains the BiHOI processing pipeline. At publication, it still labels DONIC training and evaluation code as coming soon; policy checkpoints and the BiHOI dataset are also on the release checklist.

The pipeline's own code uses CC BY-NC 4.0, with separate terms for third-party components. Model weights, MANO assets and source datasets are not bundled into that permission. Teams should check each dependency before planning commercial use.

For robotics researchers, there is enough here to inspect the capture pipeline and evaluate the method's assumptions. There is not yet a complete downloadable package for reproducing the controller from scratch.

The next useful evidence is a reproducible policy release followed by physical tests on held-out objects. Until then, DONIC offers a specific advance in learning coordinated contact, with a clear gap between simulation progress and a deployable robot skill.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn research and complex engineering into clear launch stories that developers and buyers can understand and share.

Book a 30-minute call.