Nvidia introduced GR00T 2 at the Conference on Robot Learning (CoRL) in Austin, Texas, in a blog post published November 9. The company describes it as "an open frontier world action model" built on Nvidia Cosmos, its family of world foundation models.
GR00T 2 was trained on 100,000 hours of Nvidia's open-source human egocentric video, which shows tasks from a first-person viewpoint, alongside robot and human data. Nvidia links the human video dataset through a Hugging Face space called PhysicalAI-Human-1M-Explorer.
Nvidia reports results on 50 evaluated tasks. With no task-specific demonstrations, the model completed 40 of the 50 tasks at least once. With one demonstration per task it reached 71% average success, and with five demonstrations per task 95%. Nvidia 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.
The company paired the model with other robotics releases at CoRL, including Isaac Lab 3.0, which adds support for the Newton physics engine and reports up to 2.5x training throughput over Isaac Lab 2.3 with PhysX on a vision-based dexterous manipulation benchmark, and a developer preview of Isaac Lab-Arena 0.4. It also highlighted new benchmarks built with partners, including DexBench with RLWRLD and VABAR with UC Berkeley.
The significance for robotics teams is data cost. If one or a few demonstrations are enough to adapt a pretrained model to a new task, the expensive part of deploying a robot shifts from collecting large teleoperation datasets to pretraining on cheap human video. The success rates are Nvidia's own and were measured on its chosen task set.