NVIDIA presents GR00T 2 and Isaac Lab 3.0 at CoRL
NVIDIA introduced the GR00T 2 world action model and promoted Isaac Lab 3.0, which can run on the Newton physics engine as well as PhysX, at the Conference on Robot Learning in Austin.
On 9 November 2026 NVIDIA introduced GR00T 2 at the Conference on Robot Learning (CoRL), held this year in Austin, Texas. It is an open world action model built on its Cosmos models. The training mix covers robot data and human data, the latter including 100,000 hours of NVIDIA’s open egocentric video. According to NVIDIA, with no demonstrations for the specific task, the model succeeded at least once on 40 of the 50 tasks it was tested on. It reached 71% average success over 50 tasks from one demonstration per task, and 95% from five. NVIDIA also says it transfers skills between robot bodies without robot demonstrations. Partners named as testing it include Foxconn, for server manufacturing, and Siemens.
Isaac Lab 3.0 shared the stage, although the v3.0.0 tag was published on GitHub on 20 October. Its release notes describe one task API across several physics and rendering back ends. Newton, a Warp-based engine with MuJoCo-Warp and other solvers, can now run tasks without Isaac Sim, which becomes optional. PhysX remains available. The release targets Python 3.12, PyTorch 2.11 and CUDA 13.0. On a vision-based dexterous manipulation benchmark, NVIDIA reports that Isaac Lab 3.0 on Newton trains up to 2.5 times faster than Isaac Lab 2.3 on PhysX. It also released a 0.4 developer preview of Isaac Lab-Arena, a framework for running policy evaluations across many scene variations in parallel.
Why it matters
Isaac Lab without a full Isaac Sim install lowers the cost of trying it, and the Newton back end brings deformable objects such as cables and cloth into the same training loop as rigid arms. The GR00T 2 figures are NVIDIA’s own, published in a launch post. If you plan to post-train it, look for the model card, licence and benchmark details before committing. Moving an Isaac Lab 2.x project to 3.0 is a migration, so budget time for it.