DEXOP

MIT's passive hand exoskeleton that lets a human's own hand directly drive a robot hand to collect vision and touch data.

DEXOP is a dexterous-hand data-collection device made public in September 2025 by the MIT Improbable AI Lab (Haozhi Qi, Pulkit Agrawal, and others). The authors propose a collection paradigm called perioperation: record real human manipulation while keeping the data as directly transferable to a robot as possible. DEXOP is a passive hand exoskeleton: a person's fingers are mechanically linked to the fingers of a robot hand, so whatever the person's hand does, the robot hand mirrors the same pose; the robot hand itself carries cameras and tactile sensors, so the data collected is the robot's own vision and touch data from the start. Compared with teleoperation, the operator feels contact force directly through their fingers, making the motion more natural, faster, and more accurate. The paper reports that, measured per unit of collection time, policies trained on DEXOP data clearly outperform those trained on teleoperation data.

ExampleAn operator wearing DEXOP performs contact-rich dexterous tasks while the cameras and tactile sensors on the robot hand record in sync; this data is used directly to train a policy for that same robot hand.

Also called

DEXOP: A Device for Robotic Transfer of Dexterous Human Manipulation

Related

Exoskeleton · Tactile Data · Robot-free (Embodiment-free) Data Collection · DexUMI · Teleoperation · Contact-rich Manipulation

Sources

DEXOP 项目主页 (Chinese)
DEXOP (arXiv)

As of

2025-09