Overview
- Nvidia announced GR00T 2 and Isaac Lab 3.0 at the Conference on Robot Learning in Austin on Monday, presenting them together as a connected robot‑training stack.
- GR00T 2 is a world‑action foundation model that Nvidia says learned from 402,730 hours of first‑person human video plus handheld and teleoperation data and that Nvidia will publish 100,000 hours as Human‑1M‑100K under a permissive license.
- In company tests on a preselected set of 50 manipulation tasks, Nvidia reported GR00T 2 completed 40 tasks at least once with no task demos, averaged 71% success with one demonstration, and averaged 95% success with five demonstrations after brief fine‑tuning; these results are from lab and simulation evaluations and not independent factory trials.
- Isaac Lab 3.0 is a modular training and deployment framework that lets users switch physics engines and renderers, boosts training throughput in Nvidia benchmarks, and includes AlphaNIST results showing one multitask policy can learn many assembly tasks far more GPU‑efficiently than per‑task policies.
- Nvidia named early industrial testers including Foxconn, Siemens and Nexuni, said code and dataset releases are planned for December, and cautioned that broad real‑world, factory‑scale validation and long‑term robustness remain to be demonstrated.