NVIDIA launches GR00T 2 robot model at CoRL with 95 percent claim on fifty tasks

Robot learning has spent the past several years stuck on a deceptively simple problem: there is not enough data. Language models could be trained on the open internet, but manipulation skills have to come from physical machines, and every hour of robot teleoperation is expensive, slow and specific to one arm. That constraint has shaped the entire research agenda around foundation models for robotics, pushing labs toward simulators, synthetic demonstrations and whatever human video can be scraped, licensed or recorded.

The Conference on Robot Learning, held this year in Austin, Texas from November 9 to 12, has become the annual checkpoint for that work. It is where academic groups and industrial research labs compare notes on embodiment, generalization and the awkward gap between benchmark scores and factory floors. This year the venue was dominated by one announcement from NVIDIA, which used the first day of the conference to launch its second generation robot foundation model alongside a new release of its simulation framework.

NVIDIA's GEAR lab, the research group behind the company's embodied AI efforts, introduced GR00T 2 as an open frontier world action model built on NVIDIA Cosmos. The pitch is not simply that the model performs well, but that it learns from human demonstration rather than robot demonstration, and that a skill learned on one body can be moved to another. NVIDIA's blog, published November 9, described the release as part of a broader push that also includes Isaac Lab 3.0 and a set of open robotics benchmarks.

The timing matters because the robotics industry is converging on a common architecture. Vision-language-action models map an image and an instruction directly to motor commands. World action models add a prediction step, generating the future view the robot expects to see and the action that produces it. GR00T 2 belongs to the second camp, and much of the argument around it turns on whether that extra machinery buys real generalization or just a more elaborate way to overfit a benchmark.

Key Facts

The technical specification is concrete. Unite.AI reported on November 9 that GR00T 2 is a 2 billion parameter transformer built on Cosmos-Predict2.5-2B, combining recent camera views, robot state and a written instruction to produce expected future views plus the next action sequence. A single pretrained checkpoint covers parallel-gripper arms, dexterous hands and a humanoid, which is the cross-embodiment claim in its most literal form.

Training data volumes are large by robotics standards. The same Unite.AI report put pretraining at 402,730 hours of first-person human video, 10,184 hours of handheld-gripper demonstrations across 2,897 tasks, and 23,500 hours of robot teleoperation. NVIDIA released 100,000 hours of that human data as Human-1M-100K under a CC BY 4.0 license, with video, language annotations and head and hand motion, allowing both commercial and noncommercial use.

The headline numbers are the ones NVIDIA is promoting hardest. The Robot Report reported on November 9 that GR00T 2 completed 40 of 50 evaluated tasks at least once without any task-specific demonstrations, reached 71 percent average success across those 50 tasks with one demonstration per task, and rose to 95 percent with five. Those five demonstrations supply under three minutes of robot data per task, followed by roughly a 30 minute adapter fine-tune on a single RTX 5090.

Partner activity was announced in the same breath. Nexuni is testing the model for robotic laundry handling, Foxconn is applying it to factory automation for server manufacturing, and Siemens is evaluating it across industrial use cases. MicroAGI post-trained GR00T 2 with reinforcement learning to swap Ethernet cables using a YAM robotic arm, generating and calibrating its simulation environments inside Isaac Lab 3.0 and Newton.

Isaac Lab 3.0 is the infrastructure half of the release. The Robot Report noted on November 9 that it is available on GitHub and lets researchers set up environments with simulation-ready assets, switch between the PhysX and Newton physics solvers, and choose renderers for perception training. In a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton reached up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX. Unite.AI reported that the framework has been generally available since October 20, 2026.

Analysis

The most interesting line in the release is not 95 percent. It is 40 of 50. Completing a task at least once, with no task-specific demonstration, is a qualitative capability claim rather than a reliability claim, and NVIDIA leans on it to argue that the model arrives with usable priors. The company also says GR00T 2 ranks in the top three across all vision-language-action and world action model categories on RoboCasa365, RoboLab, RoboDojo and MolmoSpaces, with multiple leaderboard wins.

The caveats are equally specific. Tech Times reported on November 10 that NVIDIA did not name the 50 tasks, disclose how many rollouts produced each average, or state a license for the weights. Unite.AI's technical read of the whitepaper adds the sharper point: the evaluated tasks were selected among those the model can learn with five episodes at a success rate of at least 80 percent, so the 95 percent average is high partly by construction. Five demonstrations that work because the task was chosen to work with five demonstrations is a different claim from five demonstrations that work generally.

Independent benchmarking is far less flattering. Tech Times reported on November 10 that the RoboDojo paper, with 44 authors and 30 tested policies, found a best simulation score of 8.80 percent and a best real-world score of 12.8 percent. Human experts scored 76.03 percent in simulation and 100 percent on real robots in the same comparison, and Stanford's 2026 AI Index, cited in the same coverage, contrasted 89.4 percent on RLBench with 12.4 percent full-task success on BEHAVIOR-1K. What this really means is that benchmark saturation in robotics is currently a measurement problem as much as a capability problem, and a vendor average without rollout counts cannot be compared to a third-party average that publishes them.

There is also an efficiency question. Tech Times reported that NVIDIA's glossary describes GR00T 2 as built on DreamZero, a February 2026 video diffusion model with 14 billion parameters, and that a March robustness study found world action models at least 4.8 times slower per inference step than the pi0.5 vision-language-action model. A model that needs a 30 minute fine-tune and predicts future frames before acting is plausible for a factory cell with a fixed camera. It is a harder sell for a mobile humanoid that has to react in real time.

Why It Matters

The bigger picture here is that NVIDIA is trying to become the default substrate for embodied AI in the same way it became the default substrate for training large language models. GR00T 2, Isaac Lab 3.0, Newton, the Video to Data workflow that converts human video into simulated robot demonstrations, and the EgoHand model for 3D hand motion reconstruction are not a product line so much as a stack. EgoHand, built with Noitom Robotics data, passed 40,000 Hugging Face downloads in two weeks, which is a signal about where developers are actually spending attention.

The openness claim is the part worth watching. NVIDIA markets GR00T 2 as open, yet Unite.AI reported on November 9 that the research page says the code will be available in December. Open weights, open code and open training data are three different commitments, and only the human data has a clear license attached so far. For Foxconn and Siemens, the license matters more than the benchmark, because factory deployments require legal clarity about derived models and redistribution.

The competitive context is a field of well funded rivals. Skild AI used the same conference to announce STEEL, a rigid-body physics solver built on Newton, and Flexion demonstrated a Reflect model combining reasoning with whole-body manipulation. NVIDIA's advantage is that it supplies the simulation, the solvers, the benchmarks and increasingly the policy, which means its success metrics are also its marketing.

Next Up

The immediate test is December, when the code is due to appear alongside the open Video to Data workflow and its accompanying research challenge. The more meaningful test is whether Nexuni's laundry handling, Foxconn's server manufacturing line and Siemens' industrial evaluations produce published results with named tasks and rollout counts. Those deployments are the only place where a 95 percent average meets a real defect rate.

Until then, the honest reading of GR00T 2 is that it is a genuine technical advance wrapped in a benchmark strategy that NVIDIA controls end to end. The model's pretraining scale and cross-embodiment transfer are real and documented, and the 30 minute adaptation budget on one consumer GPU is a meaningful practical detail. The 95 percent figure is the part that should travel with an asterisk, because the company that selected the tasks also reported the score.