NVIDIA unveils robot simulation and GR00T 2 at CoRL

Summary

NVIDIA unveiled GR00T 2 at CoRL 2026. This open-world action model, trained on 100,000 hours of human-centric data, was presented alongside Isaac Lab 3.0 and new physics integrations. NVIDIA is positioning simulation as the primary training ground for agile robot manipulation.

Why it matters

In robotics, manufacturing, and logistics, economic value is driven more by fine manipulation (cable swapping, gear alignment, millimeter-scale assembly) than by mere mobility. By offering a foundation model, an open simulation stack, and a broad partner list together, NVIDIA is creating a full-stack gravity well. This could lock the robotics research community into its ecosystem for years.

Highlights

• GR00T 2 achieved an average success rate of 71% across 50 tasks with a single demonstration per task. This rate rose to 95% when the number of demonstrations increased to five, a notable result for few-shot manipulation.
• Isaac Lab 3.0, featuring the Newton physics solver, improved training efficiency by up to 2.5x compared to the previous PhysX-based version in agile manipulation benchmarks.
• AlphaNIST achieved over 98% training success on all 20 tasks in the NIST Assembly Task Board using a single unified policy and only a binary success signal.

Critical take

This is a corporate NVIDIA blog post, not independent news. The headline claiming success 'at least once in 40 of 50 tasks' sets a deliberately low bar. Claims of 'top 3' benchmark performance rely on a small set of leaderboards with limited third-party verification. Furthermore, most named partners are still in the 'evaluation' or 'testing' phase, not yet at the deployment stage. The gap between a demo at CoRL and reality on the factory floor remains vast.