NVIDIA Unveils GR00T 2 Robot Model

What happened

NVIDIA introduced GR00T 2 on Nov 9 at the Conference on Robot Learning (CoRL) in Austin, Texas, alongside Isaac Lab 3.0, a developer preview of Isaac Lab-Arena 0.4, and two new robot evaluation benchmarks, DexBench and VABAR.

GR00T 2 is what NVIDIA calls an open frontier world action model, built on its Cosmos world models. It predicts the next camera view and the robot's next actions together. Training data includes robot operation data and first-person human video, of which 100,000 hours are open source. The model can also learn from human demonstrations and transfer skills across different robot bodies.

In NVIDIA's tests, GR00T 2 completed at least once 40 of 50 dexterous manipulation tasks with no demonstrations, and averaged 71% success with one demonstration per task and 95% with five. The model's paper discloses that the 50 tasks were chosen among those learnable with five demonstrations, a condition behind the 95% figure. Isaac Lab 3.0 adds the Newton physics engine and reaches up to 2.5 times the training throughput of version 2.3 on a dexterous manipulation benchmark. Nexuni, Foxconn and Siemens are already testing GR00T 2; model code is due in December, while the weights release date and license have not been announced.

Key facts

1. Where and when: Nov 9, at the Conference on Robot Learning (CoRL) in Austin, Texas
2. What it is: An open world action model built on Cosmos world models; it predicts future video and robot actions together
3. Training data: Robot operation data and first-person human video, of which 100,000 hours are open source
4. Test results: Across 50 dexterous tasks: at least 40 completed once with no demonstrations; 71% average with one demonstration per task and 95% with five (tasks were chosen among those learnable with five demonstrations, per the paper)
5. Isaac Lab 3.0: Open-source robot learning framework with the Newton physics engine, up to 2.5x the training throughput of version 2.3; DexBench and VABAR are new evaluation benchmarks
6. Openness: Model code is due in December; the weights release date and license have not been announced
7. Early users: Nexuni for robotic laundry handling, Foxconn for server-manufacturing automation, Siemens for industrial evaluation