NVIDIA releases GR00T 2 open world action model and Isaac Lab 3.0 at CoRL
NVIDIA says GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations; the results are NVIDIA's own and not yet independently tested.
Why it matters
NVIDIA is pitching open world action models as the base layer for robot manipulation, the step after locomotion demos. GR00T 2 is built on Cosmos and trained on robot and human data including 100,000 hours of egocentric video, and NVIDIA reports 71% average success across 50 tasks with one demonstration per task, rising to 95% with five. Isaac Lab 3.0 and new open benchmarks were released alongside, with partners including Bosch Research, Siemens, Skild AI and UC Berkeley. Watch for independent results on the new benchmarks and the CoRL best-paper decision, where two NVIDIA papers are finalists.
Key facts
- GR00T 2 is built on NVIDIA Cosmos
- Trained with 100,000 hours of human egocentric data
- 40 of 50 tasks completed at least once with no task demos
- 71% average success with one demo per task; 95% with five
- Isaac Lab 3.0 released at CoRL in Austin, Texas
- Two NVIDIA papers among six CoRL Best Paper finalists