NYU's Steertuning Lets a Robot Learn a Chore From 5 Minutes of Human Demo

CyberRobooo · x · 2026-11-11

A look at Steertuning from Lerrel Pinto's group at NYU:

A person wipes a kitchen table five times in front of a camera (5 minutes); the robot then wipes it itself.
With just 5 human demos and under 5 minutes, it reaches 71% success across 12 tasks in 12 new environments, including kitchens in real homes.
The hardware is a wheeled semi-humanoid: a YOR mobile base with two Wuji dexterous hands, learning directly from human hand motion.
Underneath is a 400M-parameter diffusion policy pretrained on 110 hours of first-person human video; hands are erased and each scene becomes 5,000 3D points, transferring human motion to robot hands.
A new task takes 2 minutes of finetuning; at deployment the robot picks the demo closest to the scene and uses that trajectory to steer action generation, with Gemini ER 1.6 planning higher-level steps.
The author notes 71% is still a research number, but learning from a few demos is how robots start giving people time back.

More from Embodied