On November 9, 2026, at CoRL in Austin, NVIDIA presented GR00T 2 alongside advances in Isaac Lab. The company says the robot model uses human and robot training data. Across 50 tasks it reports at least one success on 40 without task-specific demonstrations, average success of 71% with one demonstration and 95% with five. The first figure does not mean 80% reliability on every repetition. These are manufacturer results, not experiments reproduced by SIExpo.

From a model to a training environment

The public Isaac Lab 3.0 release is a software component distinct from the model itself. Access to training-environment code should not be described as unrestricted access to every related product's data and weights. NVIDIA's announcement emphasizes simulation and the PhysX and Newton physics engines, including more complex object interactions.

There are two separate questions: how to select an action and how to construct repeatable situations for evaluating it. A public software release lets developers inspect a specific tooling version. It does not automatically turn a laboratory-task comparison into a safety certification.

Editorial assessment: define success beforehand

For practical evaluation, we would separate at least three measures: whether a task finished, how many attempts it required and whether movement constraints were respected. This is our proposed criterion. A task succeeding once after repeated attempts establishes a capability, but does not independently describe its failure rate.

Changes between simulation and physical environments also matter. Different lighting, friction or object placement can form separate test conditions. Human demonstrations may help adaptation, but their number should accompany the reported outcome. Otherwise readers cannot distinguish immediate skill transfer from additional learning for a particular setup.

For robotics teams, the announcement is a reason to examine action models alongside evaluation infrastructure. A useful next step would be independent reproduction with documented tasks, repetitions and safety boundaries. The published percentages are a starting point for that work, rather than a promise of faultless operation in a home or production line.

Editorial verification: 2026-11-09T21:40:18.856161+00:00.