NVIDIA introduces GR00T 2 and Isaac Lab 3.0 for robot learning and simulation

#robotics #ai #machinelearning #opensource

📅 November 9, 2026 • NVIDIA blog
At the Conference on Robot Learning, CoRL 2026, in Austin, Texas, NVIDIA introduced GR00T 2, Isaac Lab 3.0 and open benchmarks for robot training. GR00T 2 is an open frontier world action model built on NVIDIA Cosmos. Its training combines robot and human data, including 100,000 hours of NVIDIA open source human egocentric data. The model transfers skills learned from human demonstrations between different robot bodies without robot demonstrations. AI agents can use its API and supporting tools to execute skills in sequence and recover from errors.

Out of the box, GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations. With one demonstration per task, it achieved 71 percent average success across 50 tasks, rising to 95 percent with five demonstrations. Completing a task at least once and average success are distinct reported measures. The model also ranks among the top three across vision language action and world action model categories for RoboCasa365, RoboLab, RoboDojo and MolmoSpaces. Nexuni is testing it for robotic laundry handling, Foxconn applies it to factory automation for server manufacturing, and Siemens is evaluating industrial use cases.

Isaac Lab 3.0, NVIDIA's open source robot-learning framework, integrates more tightly with the Newton open source physics engine to model cables, cloth and complex articulated systems such as robotic hands. In a vision-based dexterous manipulation benchmark, it achieved up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX. Thousands of simulated robots can train in parallel. AlphaNIST, built in Isaac Lab, uses a single policy for all 20 tasks on the NIST Assembly Task Board, manipulating millimeter-scale parts. It reported more than 98 percent training success against 81 percent for separate per-task policies, using roughly 39x less GPU time.

For engineering and automation teams, the practical connection is between learning individual skills, assembling them into programs and testing their behavior in simulation. UC Berkeley's Graph-as-Policy framework lets agents connect skills, test programs in NVIDIA Isaac simulation and refine them from failures while keeping programs interpretable for engineers. Success-Guided Sampling, from the University of Washington and NVIDIA, targets skills a policy is still learning and supports scaling beyond a million parallel environments. Evaluation also has defined settings: DexBench, developed by RLWRLD with NVIDIA, covers 15 tasks across four scenes spanning retail, logistics, manufacturing and hospitality. VABAR, from UC Berkeley and NVIDIA, evaluates six industrial tasks. These settings help teams relate reported results to the robot work they intend to automate.

Sources:
blogs.nvidia.com
news.berkeley.edu
research.nvidia.com