arXiv 2610.29997

DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations

RoboticsOctober 21, 2026124 upvotes

Authors: Mingze Sun, Ziang Cao, Yuan Fei, Qiankuan Tao, Zihan Wang, Tianqi Liu, Junhao Chen, Haiwen Diao +3

Overview

The authors present DONIC, a single generalist policy for bimanual dexterous reference tracking, trained jointly across operation categories and human-motion sources. They also introduce BiHOI, a multi-view dataset of 1,575 bimanual interaction sequences over 229 objects, with reconstructed interactions retargeted to robot references.

Introduction

Existing dexterous tracking models mostly handle single-hand manipulation and a narrow set of interactions, leaving common bimanual behaviors unaddressed. Contact-rich bimanual data is scarce and coordinating contacts across two hands is harder still.

Results

A single DONIC checkpoint, without finetuning, reaches 86.79% success on unseen trajectories of training objects and 57.32% on held-out objects. This surpasses state-of-the-art bimanual tracking baselines by 77.36 and 45.73 percentage points, respectively.

Conclusion

A single policy that handles heterogeneous bimanual references and generalizes to unseen objects moves robot dexterity closer to real-world two-handed tasks.

Same story, different eyes

The EngineerAI
A feedback loop that turns a robot's own successes into its teachers
What I find neat from an engineering view is that the policy recycles its own verified successful runs as demonstrations, so reinforcement learning and behavioral cloning keep feeding each other. Two hands touching one object means each hand has a different job, so rewarding each hand by its contact role while still watching the object's motion seems like a sensible way to keep them coordinated. Clip-Centered Advantages are a quiet but practical touch, since they stop easy and hard references from skewing the learning signal. How well it carries over to completely new object shapes beyond the 57% on held-out objects is still an open question.
The SkepticAI
Recycling verified successes into demos is a neat way to stretch scarce bimanual data
What caught my eye is the loop where successful executions get fed back as behavioral-cloning demonstrations, since contact-rich bimanual data is so hard to come by. The 57.32% on held-out objects is the number I'd look at most closely, because it shows how far one checkpoint carries to new shapes. It would feel even more convincing to see how results split by object type and by how long the two hands stay in contact. How the baselines were set up for the 77.36 and 45.73 point gaps is not yet known.
The FounderAI
One checkpoint for two hands opens the door to a bimanual data flywheel
Looking at this as a founder, the interesting part is that a single policy now tracks two-handed manipulation on objects it has never seen, and its own verified successes get recycled as training demonstrations. That loop could become a real moat: whoever collects the most contact-rich bimanual human data early may compound fastest. A concrete way in is to build a narrow data service, such as capturing and retargeting two-handed household or assembly tasks for one customer type, while general tracking still sits at 57% on new objects. How well it transfers to physical robot hands outside simulation isn't known yet, and that gap is probably where a specialized team can still find room.