NVIDIA's GR00T 2 Clears 40 of 50 Tasks Untrained; Licence Unstated

NVIDIA released GR00T 2, a robot foundation model it calls an open world action model, at the Conference on Robot Learning in Austin, Texas, on Nov. 9. The company said the model completed at least one trial on 40 of 50 evaluated tasks with no task-specific demonstrations.

The announcement came in an NVIDIA blog post and was covered the same day by The Robot Report. Every performance figure below comes from NVIDIA or its partners. Neither page names an independent party that ran the tests.

What NVIDIA says GR00T 2 does

GR00T 2 is built on NVIDIA Cosmos and trained on robot and human data, including 100,000 hours of NVIDIA's open-source egocentric human video. NVIDIA said it learns from human demonstrations and moves skills between different robot bodies without robot demonstrations.

With one demonstration per task, the model averaged 71% success across 50 tasks. With five, it averaged 95%.

1. 1 demonstration71 %
2. 5 demonstrations95 %

Source: NVIDIA blog post, Nov. 9, 2026 (vendor-supplied)

NVIDIA also said the model ranks among the top three in the vision-language-action and world action model categories on four benchmarks: RoboCasa365, RoboLab, RoboDojo and MolmoSpaces. The posts do not give scores or name who ran the leaderboards.

Who is testing it

The three named industrial users are testing or evaluating, not running it in production. No post announces a deployment.

Organisation	What NVIDIA says it is doing
Nexuni	Testing GR00T 2 for robotic laundry handling
Foxconn	Applying it to server manufacturing automation
Siemens	Evaluating it across industrial use cases
MicroAGI	Ethernet cable swapping on a YAM arm, using a reinforcement-learning post-trained version

Isaac Lab 3.0 and the throughput claim

NVIDIA shipped Isaac Lab 3.0, its open-source robot-learning framework, alongside the model. Users can switch between the PhysX and Newton physics solvers, and a new reinforcement post-training library targets foundation models such as GR00T.

On a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton reached up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX, according to NVIDIA. The Robot Report described that figure as reported, not independently measured.

The Robot Report also relayed a partner result: AlphaNIST, built in Isaac Lab, uses one policy for all 20 NIST Assembly Task Board tasks. It reached over 98% training success, against 81% for per-task policies, using about 39 times less GPU time. That is a training-success figure, not a result on a customer floor.

What the posts leave out

Neither the NVIDIA post nor The Robot Report states GR00T 2's parameter count, its licence, its price or a date when developers can download it. The word "open" appears in NVIDIA's description, but without licence terms it does not tell a company whether it can ship the model in a product. NVIDIA did not say which robot bodies the 50-task evaluation ran on.

Ken Goldberg, a professor of robotics at UC Berkeley who leads the VABAR and GaP projects, is quoted in the NVIDIA post on agentic robotics: "To me, agentic robotics is the most exciting technological development since the internet in 1995." The quote concerns the field, not GR00T 2 results.

Related humanoid coverage on this site includes the Boston Dynamics Atlas wheels testbed and the Agility Robotics listing plan. Nothing in the GR00T 2 posts says which humanoid makers will adopt the model. The next concrete signal is whether NVIDIA publishes a licence and a download page for the weights.