Hand over hand teaching cut dexterous robot training time in half

Proceedings of Machine Learning Research2025-10-07Full text reviewedOpen source ↗

INVESTOR TAKEAWAY

KineDex lets a person physically guide a robot hand, collecting tactile demonstrations faster and improving contact-heavy task success, but the setup is not yet scalable to two hands.

ORIGINAL PAPER

KineDex: Learning Tactile-Informed Visuomotor Policies via Kinesthetic Teaching for Dexterous Manipulation

WHAT WE KNOW

Average task success was 74.4%, a reported 57.7% improvement over the no-force-control variant.

Demonstrations were collected more than twice as fast as teleoperation in two comparison tasks, with near-100% collection success.

WATCH NEXT

Bimanual collection with one operator, demonstrations per hour at scale, policy performance after dataset growth, sensor durability, and total data cost.

EVIDENCE

What the paper supports

Supported findings

Average task success was 74.4%, a reported 57.7% improvement over the no-force-control variant.

Abstract and p. 2

Demonstrations were collected more than twice as fast as teleoperation in two comparison tasks, with near-100% collection success.

Abstract; p. 8, comparison

Limitations

Severe human-hand occlusion can break the inpainting step.

Two human hands are needed to guide one robotic hand because of thumb morphology.

The data-collection advantage is shown on a small task set, not a production fleet.

LISTED COMPANY IMPACT

Companies with a traceable link

TeslaTSLA UW

Supports Short

The need for tactile hardware and labor-intensive teaching reinforces the gap between demos and scalable deployment.

Thematic