NVIDIA launches GR00T 2 and Isaac Lab 3.0 with open benchmarks for robot learning

At the Conference on Robot Learning (CoRL 2026) in Austin, Texas, NVIDIA announced three related releases for robot learning: GR00T 2, a new foundation model for robot control; Isaac Lab 3.0, an update to its open source robot-learning framework; and a set of open benchmarks for evaluating robot policies. The details come from a dev.to report on NVIDIA's blog post and related pages from NVIDIA Research and UC Berkeley.

GR00T 2 learns from human video

GR00T 2 is described by NVIDIA as an open "world action model" built on the company's Cosmos platform. Its training mixes robot data with human recordings, including 100,000 hours of open source egocentric human video that NVIDIA collected. The headline capability is transfer: skills learned from human demonstrations can move between different robot bodies without demonstrations collected on each robot. NVIDIA also says AI agents can call the model through an API and supporting tools to run skills in sequence and recover when a step fails.

The reported numbers

According to the dev.to report, GR00T 2 completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations. A single demonstration per task lifted average success across the 50 tasks to 71 percent, and five demonstrations brought it to 95 percent. NVIDIA explicitly distinguishes between completing a task at least once and average success rate, since they measure different things. On public comparisons, the model ranks among the top three across vision language action and world action model categories for RoboCasa365, RoboLab, RoboDojo and MolmoSpaces.

Several companies are already testing it: Nexuni for robotic laundry handling, Foxconn for factory automation in server manufacturing, and Siemens for a range of industrial use cases.

Isaac Lab 3.0 and the Newton physics engine

Isaac Lab 3.0 integrates more tightly with Newton, an open source physics engine that can model cables, cloth and complex articulated systems such as robotic hands. In a vision-based dexterous manipulation benchmark, the new release reached up to 2.5 times the training throughput of Isaac Lab 2.3 running with PhysX, and it supports thousands of simulated robots training in parallel.

One showcase project, AlphaNIST, trains a single policy to handle all 20 tasks on the NIST Assembly Task Board, manipulating parts at millimeter scale. NVIDIA reports over 98 percent training success with this unified policy, compared with 81 percent for separate per-task policies, while consuming roughly 39 times less GPU time.

Open benchmarks and evaluation settings

Alongside the model releases, NVIDIA and its partners published evaluation settings intended to make reported results comparable across labs. 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.

Two research frameworks address the gap between learning individual skills and shipping them as programs. UC Berkeley's Graph-as-Policy lets agents connect skills, test the resulting programs in Isaac simulation and refine them from failures while keeping the programs interpretable for engineers. Success-Guided Sampling, from the University of Washington and NVIDIA, concentrates training on skills a policy has not yet mastered and reportedly scales beyond a million parallel environments.

Why it matters

Robot learning has historically been fragmented: policies were trained per robot and per task, and results were difficult to compare across research groups. GR00T 2's claim of transferring skills from human video to multiple robot bodies attacks the data-collection bottleneck directly, since recording humans is far cheaper than teleoperating robots for demonstrations. The demonstration-count curve, 71 percent success with one example and 95 percent with five, points toward a practical workflow where teams adapt a general model with a handful of demonstrations instead of running full training pipelines.

Isaac Lab 3.0's throughput gains and Newton's deformable-object simulation lower the cost of the other half of the loop, letting engineers assemble skills into programs and test their behavior in simulation before anything touches a factory floor. With Foxconn and Siemens already evaluating the stack for industrial work, the announcement reads less like a research demo and more like an attempt to make generalist robot policies a standard component in automation engineering.