Morning Edition · Friday, October 23, 2026Published at 2:19 AM EDT · New York

Real2Sem2Real Turns One Human Video Into Thousands of Robot Demonstrations That Match Teleoperation Data

The pipeline generates more than 2,000 training episodes in about two hours from a single monocular RGB recording, and policies trained on that data matched or beat teleoperation baselines on two real robots and four tasks.

A paper posted to arXiv on October 20, Real2Sem2Real, addresses the most expensive input in robot learning: demonstration data. The pipeline starts from one monocular RGB video of a person doing a task. It produces thousands of robot demons…