NVIDIA unveils GR00T 2 and Isaac Lab 3.0 at CoRL

NVIDIA unveiled GR00T 2 and Isaac Lab 3.0 on November 9, 2026, at the Conference on Robot Learning (CoRL) in Austin, Texas. According to The Robot Report, GR00T 2 was trained on diverse data from robots and people, including 100,000 hours of NVIDIA open-source first-person-view data. It reportedly completed at least once 40 of 50 evaluated tasks without task-specific demonstration data.

NVIDIA’s announcement presented a benchmark stating that the model completed at least once 40 of 50 tasks without task-specific demonstrations, putting the evaluation conditions used to compare model capabilities in focus. The precision-manipulation examples cited in NVIDIA’s official blog show that simulation training is being applied to tasks beyond walking, but they do not provide figures that validate performance on physical robots.

To assess the scale of the announcement, GR00T 2’s “100,000 hours” and “40 of 50 tasks” figures indicate data volume and evaluation results, but do not, on their own, show the model’s success rate in real-world settings or the breadth of tasks it can perform. An evaluation that counts a task as completed at least once is a different measure from consistency or safety in repeated operations. These figures should therefore be read as limited indicators for evaluating the capabilities NVIDIA claims, not as proof of established generality. A notable feature of the announcement is that it presented a model and a training framework at the same event. However, the information provided does not confirm that GR00T 2 was trained or evaluated on Isaac Lab 3.0, or that the two operate as an integrated product. The NVIDIA official blog’s examples of manipulating gears, cables and small parts provide context on simulation being used for precision manipulation, but there is no basis for linking these examples to the same tests as GR00T 2’s 40-task evaluation. An article by Nikkei XTech on humanoid control technology using world models, covered by this newspaper, focused on reinforcement learning using world models and Sim-to-Real for humanoids. The current announcement also concerns both models and training environments, but the earlier report covered whole-body control, while the examples in NVIDIA’s current official blog involve precision object manipulation. It is important not to conflate the evaluation targets or stages of physical-robot deployment. Meanwhile, NVIDIA’s demonstration of manually guiding a robotic arm with Holoscan 5 covered an operation in which a person physically moved the arm. That has a different role from model training and simulation in the current announcement. The broader industry implication is the need to evaluate training data, simulation environments and physical-robot testing separately. Even when data volume is reported, it remains unclear whether skills transfer across different robots or whether skills acquired in simulation can be reproduced on physical robots. Questions to examine include the specific changes in Isaac Lab 3.0, the 40 tasks and their success criteria, repeated results for each task, conditions for physical-robot testing, and the range of supported robots. The cited materials do not provide these details.

NVIDIA’s official blog cites aligning gears, connecting cables and assembling small parts as examples of robot training using simulation. It says simulation is being used not only for walking, running and jumping but also for tasks requiring finer manipulation. The available excerpt from the official blog does not describe changes in Isaac Lab 3.0 or technical integration specifications with GR00T 2. The fact that both were announced should therefore be treated separately from the uses of simulation described in the blog.