Nvidia Releases GR00T 2, an Open World Action Model for Teaching Robots Dexterous Tasks

The model belongs to a class Nvidia calls world action models, which learn both how an environment is likely to change and which actions a robot can take to bring that change about. GR00T 2 is built on Nvidia's Cosmos world models and was trained on a mix of robot and human data, including 100,000 hours of human egocentric video that Nvidia has released as open source.

The release matters because the hard problem in robotics has moved from movement to manipulation. Walking and rolling robots have been trained in simulation for years. Tasks such as aligning gears, swapping cables or handling cloth need far finer control, and a general model that learns them from little data would cut the cost of putting robots on factory and warehouse floors.

Performance Claims and Early Users

Nvidia said GR00T 2 completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations. With one demonstration per task, it reached 71% average success across the 50 tasks, rising to 95% with five. The company said the model ranks among the top three across vision language action and world action model categories on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks. These are Nvidia's own figures and have not been independently verified.

AI agents can also call GR00T 2 through an application programming interface to chain skills together and recover from errors, Nvidia said. Foxconn is applying the model to automating server manufacturing, Siemens is evaluating it for industrial uses, and Nexuni is testing it for robotic laundry handling. MicroAGI has post-trained it with reinforcement learning to swap Ethernet cables using a YAM robotic arm, a workflow it will demonstrate at the conference.

Simulation Tools and Benchmarks

Alongside the model, Nvidia released Isaac Lab 3.0, the latest version of its open-source robot-learning framework. It lets researchers use agents to set up training environments and switch between the PhysX and Newton physics engines. Tighter integration with Newton, an open-source engine, means Isaac Lab can now model deformable objects such as cables and cloth, as well as robotic hands with many interacting joints. On a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton delivered up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX, the company said.

Skild AI announced STEEL, a rigid-body physics solver built on Newton to speed up robot training. Nvidia also previewed Isaac Lab-Arena 0.4, which runs policies across thousands of variations in parallel to find the conditions that make them fail. Korean startup RLWRLD and Nvidia released DexBench, which tests dexterity across 15 tasks in retail, logistics, manufacturing and hospitality scenes. UC Berkeley and Nvidia released VABAR, a benchmark covering six industrial tasks.

For data collection, Nvidia released an open Video to Data workflow that turns human video into simulated robot demonstrations. Its EgoHand model, which reconstructs 3D hand motion, has passed 40,000 downloads on Hugging Face in the two weeks since launch, the company said.

Nvidia is also backing what it calls agentic robotics, in which frontier AI agents write code that robots can run and that engineers can read. UC Berkeley's Graph-as-Policy method uses agents to link individual skills into robot programs, then tests and refines them in Isaac simulation before deployment. Ambi Robotics, Bosch Research and Siemens are exploring it for industrial work. "To me, agentic robotics is the most exciting technological development since the internet in 1995," said Ken Goldberg, a robotics professor at UC Berkeley.

Two Nvidia papers are among six finalists for CoRL's Best Paper Award: GRAIL, a pipeline that generates 3D assets and simulator-ready scenes for robot learning, and RoboTTT, a training recipe that Nvidia says scales a robot policy's visual and motor context by three orders of magnitude without increasing inference latency. The winners will be announced at the conference this week.