NVIDIA showed GR00T 2, a world‑action model that learned dexterous manipulation from human video and succeeded on 40 of 50 tasks without robot‑specific demos. The model also reached 71% success with a single demonstration per task and 95% with five. At the same event, NVIDIA released Isaac Lab 3.0, which now runs with the Newton physics engine to simulate deformable objects and claimed up to 2.5× faster training than the previous version.
Partners such as Foxconn, Siemens and the startup MicroAGI are testing GR00T 2 for factory automation and cable‑swapping, while developers can use the new Isaac Lab‑Arena preview to evaluate robot policies across thousands of variations.
Why it matters
Researchers can now train robots on fine‑grained assembly work in simulation, shortening the path to real‑world industrial deployment.
