NVIDIA announces GR00T 2, an open world action model for humanoid robots

NVIDIA announced GR00T 2, an open world action model for humanoid robots, at the CoRL 2026 robot learning conference in Austin, Texas, on 9 November.

The model builds on NVIDIA's Cosmos world foundation model and learns from human and robot data to grasp and manipulate objects, the company said. It trained on 100,000 hours of first-person human video that NVIDIA had released, among other data, and can apply skills learned from human hand movements to robot arms, five-fingered hands and humanoids.

NVIDIA said GR00T 2 ranks in the top three on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks for vision-language-action and world action models. It also links to AI agents through an API, so multi-step tasks can be chained with error recovery.

The company also introduced Isaac Lab 3.0, an open source robot learning framework with a choice of its PhysX engine or the open source Newton engine, and Video to Data, a workflow that turns videos of people working into robot demonstration data for simulation.