NVIDIA announced GR00T 2 on Monday at the Conference on Robot Learning in Austin, and released 100,000 hours of first-person human video to go with it. The model is what NVIDIA calls an open world action model: before it moves, it predicts how the action will change what the robot sees. It was trained on robot data and human video, which is where the hand skills come from.
The numbers NVIDIA published are strong and carry a caveat the company states itself. GR00T 2 completed 40 of 50 tasks at least once with no task-specific training. Shown a task once, it succeeded 71% of the time on average. Shown it five times, 95%. NVIDIA’s own paper notes those 50 tasks were selected from ones the model could already learn well, so read the figures as a ceiling rather than a field average.
Alongside it came Isaac Lab 3.0, also open source, which NVIDIA says trains up to 2.5 times faster on one benchmark. Foxconn, Siemens and Nexuni are testing GR00T 2; none has committed to using it in production. The release matters less for what it does today than for what it removes: the training footage is the expensive part of robot learning, and NVIDIA just gave away 100,000 hours of it.
Sources: The Next Web.
Published 10 November 2026. Every figure above was checked against the vendor page or the reporting linked here on the same day.