NVIDIA releases GR00T 2 open world action model and Isaac Lab 3.0 with Newton integration at CoRL 2026

At the CoRL robot learning conference in Austin, NVIDIA announced GR00T 2, an open robot model built on its Cosmos world model and trained on robot data plus human data, including 100,000 hours of first-person video. It also released Isaac Lab 3.0, a simulation toolkit that can switch between the PhysX and Newton physics engines and models soft objects such as cables and cloth. NVIDIA's own figures: GR00T 2 completed 40 of 50 tasks at least once with no task-specific demonstrations, reached 71% average success with one demonstration per task and 95% with five, and Isaac Lab 3.0 with Newton ran up to 2.5x faster in training than Isaac Lab 2.3 on one benchmark. The source is NVIDIA's own newsroom post, and no linked paper, model card or independent test was cited.

Robotics researchers and developers get an open model and a free-to-use simulation stack from the dominant AI chip supplier, which could lower the cost of teaching robots fiddly manipulation tasks. The main effect is likely to be on who builds robot software on NVIDIA's tools, which strengthens NVIDIA's position in a field where rivals such as Skild AI are also building their own physics tools. Anyone considering factory deployment should treat this as a research and prototyping tool for now. Nothing here shows robots working reliably in industrial settings.

The line about robots moving from lab demonstrations to industrial deployment is framing, not a result. "Completed 40 of 50 tasks at least once" is a low bar, as one success in any number of tries counts. The 71% and 95% success rates and the 2.5x speed-up come from NVIDIA's chosen benchmarks, with no detail on task difficulty, hardware or what the comparison baselines were. The long partner list shows interest, not proof that any of them have deployed the technology.

1. 01Whether NVIDIA publishes weights, a technical paper and evaluation code for GR00T 2 that outsiders can run, and whether independent labs reproduce the 40 of 50, 71% and 95% figures in the next few months.
2. 02Whether the 2.5x training speed-up for Isaac Lab 3.0 with Newton holds on other people's workloads, shown in community benchmarks or GitHub issues after release.
3. 03Named partners such as Foxconn or Siemens reporting real production use with measured results, not pilots or demos, over the next two quarters.