Nvidia's GR00T 2 picks up new robot skills after one demonstration
Nvidia unveiled GR00T 2 in Austin, a robot model of about 2 billion parameters that reaches 71% average success on 50 tasks after a single demonstration.

Nvidia presented GR00T 2 on November 9 at the Conference on Robot Learning in Austin, Texas, together with Isaac Lab 3.0. The model has about 2 billion parameters and is built on the Cosmos platform. Rather than only choosing actions, it first predicts what its cameras will see next and then plans the movements that follow. Much of its training came from first-person footage of people at work, adding up to 100,000 hours.
According to Nvidia, the model completed 40 of 50 tested tasks at least once with no demonstrations. With one demonstration per task it averaged 71% success, and with five it reached 95%. Nvidia also says it ranks in the top three on four benchmarks, but these are its own results on tests it chose. Isaac Lab 3.0, which uses the Newton physics engine, trains dexterous skills up to 2.5 times faster than version 2.3. Two new industrial benchmarks, DexBench and VABAR, were released too, and Siemens, Bosch and Ambi are named as collaborators.
Why it matters
Collecting demonstrations by remote control is slow and expensive, so a model that learns from human video and a single example could cut the cost of teaching robots new jobs. The model weights are expected in December, which would let outside teams test the claims. Nvidia has not yet shared independent results or shown how the model performs beyond the benchmarks it picked.