Paper review · October 23, 2026
DONIC: General-Purpose Bimanual Dexterous Manipulation Tracking Control Learned from Human Demonstrations
DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations
By combining the bimanual manipulation dataset BiHOI, object-centric rewards and clip-level Advantage correction, the authors trained a policy that tracks a wide range of bimanual manipulation with a single checkpoint.
Problem
Existing tracking policies that have robot hands imitate human demonstrations have mostly been limited to single-hand manipulation. Data in which both hands stay in contact is scarce. In addition, the roles of the two hands and the timing of contact keep changing, which makes it hard to learn with a single policy.
Method
The authors built the BiHOI dataset (1,575 sequences, 229 objects, 14.77 hours) by recording bimanual manipulation with 32 calibrated RGB cameras. They reconstructed hands and objects in 3D and then retargeted them to Wuji robot hand motions. Following DexTrack, they alternate between PPO and behavior cloning (BC) training and reuse verified successful runs as new demonstration data. The reward centers on tracking object position, rotation and velocity, and each hand receives a contact reward only when that hand should be in contact in the reference motion. CCA (Clip-Centered Advantages) subtracts the per-clip mean Advantage to remove bias caused by differences in clip difficulty.
Results
- The authors reported that, without fine-tuning, a single checkpoint achieved an 86.79% success rate on new trajectories with training objects and 57.32% on unseen objects.
- Its success rates were 77.36 and 45.73 percentage points higher, respectively, than the stronger of the DexTrack and Dexplore baselines adapted for two hands.
- In BiHOI, 65.91% of frames with at least one hand in contact had both hands in contact. This is 27.34 percentage points higher than H2O (38.57%), the highest among the public datasets analyzed.
Limitations
In the provided text, evaluation was done only in simulation on public data splits, and no real-robot experiments are confirmed. Only one robot hand, Wuji, was used. The success rate drops to around 57% on unseen objects, so object generalization remains limited. The comparison baselines are methods originally designed for one hand that were adapted for two hands.
Why it matters
Asymmetric bimanual tasks, such as holding something with one hand and opening it with the other in housework or assembly, are common in daily life. The work shows a way to handle them with a single policy and also releases a dataset focused on bimanual contact.
Key terms
- Reference tracking
- The task of controlling a robot so that it reproduces, inside a physics simulation, the target hand and object motions taken from human demonstrations.
- Retargeting
- The process of converting human hand motion into joint values for a robot hand with a different joint structure.
- Advantage
- In reinforcement learning, a value that shows how much better an action was than average. It sets the direction of policy updates.
- Behavior cloning (BC)
- A method that trains a model, in a supervised way, to directly imitate the observation-action pairs of successful demonstrations.
Original
Authors: Mingze Sun, Ziang Cao, Yuan Fei, Qiankuan Tao, Zihan Wang, Tianqi Liu, Junhao Chen, Haiwen Diao, et al. · Posted 2026-10-21 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.