NVIDIA Releases GR00T 2 Robot World Action Model and Isaac Lab 3.0
Posted As of
NVIDIA introduced GR00T 2, an open model based on Cosmos, timed to CoRL. All performance figures are NVIDIA's own tests, and no information on the license or weights has been confirmed.
What happened
Based on the official NVIDIA blog announcement.
- GR00T 2 is an open frontier world action model built on NVIDIA Cosmos, trained on robot and human data. This includes 100,000 hours (official) of NVIDIA's open-source first-person perspective data.
- NVIDIA claims it ranks in the top 3 across all VLA and world action model categories on the RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces benchmarks.
- A developer forum post states that the combination of Isaac Lab 3.0 and Newton delivers up to 2.5x training throughput (official) on a dexterous manipulation benchmark.
- AlphaNIST reports a simulation training success rate of over 98% (official) across 20 assembly tasks, and says it uses about 39x less GPU time (official) than individual policies.
Our take
The direct impact on local AI users is small; this is closer to news in the robotics and physical AI field. All figures come from NVIDIA's own tests. We could not confirm the model card, license, or whether weights are public from the primary page, so it is still unknown whether the model can actually be downloaded and used.
Sources and related posts
- NVIDIA Blog Primary source