CHORD: Contact-wrench guidance and 4,739-task dexterous benchmark

CHORD, short for Contact Wrench Guidance from Human Demonstration in Robotic Dexterous Manipulation, is a reinforcement-learning framework that uses object-centric forces and torques from human demonstrations to guide long-horizon, contact-rich robot manipulation. The paper was submitted to arXiv on 22 Jun 2026 (arXiv:2607.00033) and is authored by Xinghao Zhu and 19 other authors.

What is CHORD and how does it work?

CHORD represents both human and robot motions in the object-centric contact wrench space, meaning it models the forces and torques that agents can apply to the object and measures similarity by the induced instantaneous motions; this guidance is used to steer reinforcement learning for dexterous tasks. The framework reframes demonstrations as contact-wrench signals rather than joint trajectories, so reinforcement learning receives guidance tied directly to the object's interactions rather than agent-specific kinematics.

The paper frames this contact-wrench guidance as a scalability aid for contact-rich dexterous manipulation, enabling longer-horizon policies and generalization across different embodiments of the demonstrator and the robot. The authors name the approach CHORD and position it as a way to transfer human demonstration intent into robot action via forces and torques on the manipulated object.

What benchmark did the authors build and how was CHORD evaluated?

The authors created a large-scale simulation benchmark comprising 4,739 bimanual dexterous manipulation tasks, assembled from motion-capture datasets and reconstructed in-house videos, and evaluated CHORD on a subset of 1,831 benchmark tasks. On those 1,831 evaluated tasks CHORD achieved an average success rate of 82.12%, the paper reports, demonstrating the method's scalability across a wide range of contact-rich scenarios.

The benchmark is explicitly constructed to cover both rigid and articulated objects, which supports the paper's claim that contact-wrench guidance helps for varied object dynamics. The dataset origins and task count are presented as part of the contribution: 4,739 tasks in simulation and an evaluation run on 1,831 tasks drawn from that collection.

How does CHORD generalize and transfer to the real world?

CHORD generalizes beyond hand-only imitation: the authors report that the method transfers from hand-only and third-person demonstrations to whole-body manipulation, yielding a 90.77% success rate on that whole-body generalization evaluation. The learned policies also transfer to real-world hardware, with the paper stating that policies were deployed both open-loop and closed-loop on real robots.

Those claims position CHORD as a method that can bridge differing demonstration perspectives and embodiments while maintaining high task success in simulation and viability in real-world execution.

Why it matters

Representing demonstrations as contact wrenches shifts the focus from how an agent moves to what it makes the object do. That removes a layer of mismatch when human and robot kinematics differ and can make reinforcement learning guidance more directly relevant to the manipulation objective. The paper supplies concrete scale (4,739 tasks) and performance numbers (82.12% on 1,831 tasks; 90.77% for whole-body generalization) that show this representation can support both broad benchmarks and specific transfer scenarios.

If contact-wrench guidance continues to deliver those success rates across more platforms and more diverse physical setups, it would lower the barrier for using human demonstrations in dexterous robot learning and reduce the need for demonstrator-robot retargeting.

What to watch

Look for public release of the 4,739-task benchmark and code linked from the arXiv entry (arXiv:2607.00033) to enable independent replication and comparison. Also watch for follow-up work reporting more detailed real-world transfer metrics and the community using the benchmark to compare other demonstration-to-policy approaches.

References: arXiv:2607.00033, "Learning Dexterous Manipulation Using Contact Wrench Guidance From Human Demonstration", submitted 22 Jun 2026, authors led by Xinghao Zhu.