THE ROBOT●DAILY
Humanoid Robotics & Embodied AI News
← Latest briefingFrontier Tech
Frontier Tech

NVIDIA Unveils GR00T 2, a World Action Model That Learns Dexterous Skills From Human Video

The Robot Daily·2026-11-09·2 min read
NVIDIA Unveils GR00T 2, a World Action Model That Learns Dexterous Skills From Human Video

Built on Cosmos, the open model learned from 100,000 hours of first-person human video - and did 40 of 50 tasks with no demos.

NVIDIA has unveiled GR00T 2, an open world action model that teaches robots dexterous manipulation by watching people. Announced at the Conference on Robot Learning in Austin on November 9, the model is built on NVIDIA Cosmos. A world action model predicts how its actions will change what it sees before it moves - the idea being that a robot that can imagine the result of a movement can plan a delicate one.

GR00T 2 trained on a mix of robot data and human data, including 100,000 hours of first-person human video that NVIDIA has released as open source. The footage - the open-sourced subset, known as Human-1M-100K - shows people doing everyday tasks with their hands. The point is scale: instead of collecting thousands of robot demonstrations for every skill, the model learns the shape of human dexterity from video, then transfers it to robot bodies.

On NVIDIA's own tests, GR00T 2 completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations at all. With one demonstration per task, average success reached 71 percent. With five, it reached 95 percent. Those are NVIDIA's numbers, and they have not been independently verified. One caveat matters: the 50 tasks were pre-selected for being learnable from five demonstrations, so the 95 percent figure is partly by construction.

NVIDIA also says GR00T 2 ranks among the top three models across vision-language-action and world action model categories on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks, with multiple leaderboard wins. Alongside the model, NVIDIA released an open Video to Data workflow that turns human video into simulated robot demonstrations - a way to grow training data without more teleoperation.

Nexuni, Foxconn, Siemens and MicroAGI are among the first testing GR00T 2, on tasks from laundry handling to AI server manufacturing and swapping Ethernet cables. NVIDIA says it plans to make the model's code available in December, but the announcement gave no weights link, licence or release date - so for now, the model is open in name, and developers will have to wait for the details. The same CoRL announcements included Isaac Lab 3.0, an Isaac Lab-Arena 0.4 developer preview and two open industrial benchmarks.

SourcesPhoto: NVIDIA. Sources: NVIDIA