GR00T 2: 71% robot task success from one demo

What NVIDIA introduced

At the Conference on Robot Learning (CoRL) in Austin, NVIDIA introduced GR00T 2 on November 9. It is a world action model: the robot not only sees and follows instructions but anticipates how its actions will change the environment. It is built on the Cosmos world model and trained on robot and human data, including 100,000 hours of open egocentric human video (NVIDIA).

What the numbers show

• No demonstrations: completed 40 of 50 evaluated tasks at least once.
• One demonstration per task: 71% average success across 50 tasks.
• Five demonstrations: 95% success.
• Top three among VLA and world action models on RoboCasa365, RoboLab, RoboDojo and MolmoSpaces.

It learns from human demonstrations and transfers skills across robot bodies without robot demonstrations. AI agents can call it through an API to run skills in sequence and recover from errors. These are NVIDIA's own results.

Who is using it

Foxconn is applying it to server manufacturing automation, Siemens is evaluating it for industrial use and Nexuni is testing it for laundry handling. MicroAGI uses an RL post-trained GR00T 2 to swap Ethernet cables with a YAM robotic arm, building its simulation environments in Isaac Lab 3.0 and the Newton physics engine (The Robot Report).

Caveats

NVIDIA calls the model open, but the announcement gives no model size, license or download link, so it is not yet clear when or on what terms you can try it. Earlier GR00T N1.x models shipped with open weights on Hugging Face.

Training runs on Isaac Lab 3.0, which we covered earlier (NVIDIA Isaac adds reference humanoid, Isaac Lab 3.0). NVIDIA says Isaac Lab 3.0 with Newton reaches up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX on a vision-based dexterity benchmark. Fine manual work such as cables, gears and cloth has been one of robotics' hardest problems. If 71% from a single demo holds up outside NVIDIA's lab, it cuts the cost of teaching robots new tasks considerably.