Summary
NVIDIA introduced GR00T 2, Isaac Lab 3.0 and new robot-evaluation benchmarks at CoRL. The company reports gains on simulated dexterous-manipulation tasks, including faster training and success across multiple assembly tasks.
NVIDIA introduced its GR00T 2 robot model, the Isaac Lab 3.0 simulation framework and new evaluation tools at the Conference on Robot Learning (CoRL) in Austin. The announcements focus on training robots for precise manipulation—such as handling cables or assembling small parts—rather than only movement such as walking.
NVIDIA reports that GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations. In another reported test across the 50 tasks, its average success rate was 71% with one demonstration per task and 95% with five. These are company-reported results from robot-learning evaluations; the announcement does not describe them as measurements of factory performance.
A model and simulation stack for manipulation
GR00T 2 is described by NVIDIA as an open world action model built on its Cosmos platform. It was trained on robot and human data, including 100,000 hours of NVIDIA open-source, first-person human video. The company says the model can learn skills from human demonstrations and transfer them between different robot bodies without demonstrations from those robots.
A world action model aims to connect perception with action: it must interpret its surroundings, anticipate how an action could change them and carry out a sequence of movements. In practice, training and testing those skills in simulation can provide many repeatable attempts before a policy is run on a physical robot.
Isaac Lab 3.0 adds tools for setting up training environments and switching between NVIDIA’s PhysX and the open-source Newton physics engine. Newton integration lets simulations represent deformable objects, including cables and cloth, as well as articulated systems such as robotic hands. NVIDIA says that, in a vision-based dexterous-manipulation benchmark, Isaac Lab 3.0 with Newton achieved up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX. Throughput here refers to the rate of training in that benchmark, not a general measure of robot performance.
One example is AlphaNIST, a policy trained in Isaac Lab to perform all 20 tasks on the NIST Assembly Task Board, which uses millimetre-scale parts and a yes-or-no success signal. NVIDIA reports more than 98% training success for the single policy, compared with 81% for separate policies trained for each task, using roughly 39 times less GPU time. The figures describe simulation training results; transferring a learned policy to hardware also depends on how closely the simulation represents real objects and conditions.
Testing whether skills generalise
A robot that succeeds once may still fail when an object moves or its physical properties change. NVIDIA’s open-source Isaac Lab-Arena framework is intended to evaluate policies across many variations in parallel and analyse the conditions associated with failures. Its developer preview, version 0.4, was announced for CoRL.
Two new benchmarks use the framework to test industrial manipulation. DexBench, developed by RLWRLD with NVIDIA, covers 15 tasks in four retail, logistics, manufacturing and hospitality scenes, with variations in objects, spatial arrangements and physics. VABAR, developed by UC Berkeley and NVIDIA, evaluates six industrial tasks using standardised variations, difficulty tiers and success measures.
The announcement also describes agent-based workflows for combining individual skills into longer robot programs. UC Berkeley’s Graph-as-Policy framework uses agents to test and refine programs in simulation after failures, while keeping the resulting programs interpretable for engineers to review. The approach addresses a practical step between learning a single movement and coordinating a robot through a multistep job. Its value in deployment will depend on how reliably these simulated skills and programs transfer to the intended physical setting.