NVIDIA TAUGHT ROBOTS TO LEARN BY WATCHING HUMANS — 95% SUCCESS AFTER FIVE DEMOS

Nvidia unveiled GR00T 2 at the Conference on Robot Learning in Austin on November 9 — an open frontier world action model built on Cosmos that predicts future video and robot actions together, trained on 402,730 hours of first-person human video plus robot teleoperation. It completed 40 of 50 evaluated tasks at least once with no demonstrations, averaging 71% success with one demonstration per task and 95% with five.

The caveats are Nvidia's own numbers on unnamed tasks: code arrives in December, and no license or weights release date has been announced. Foxconn, Siemens and Nexuni are testing it, Nvidia shipped Isaac Lab 3.0, and 100,000 hours of the human video is now public as Human-1M-100K under a Creative Commons license.