NVIDIA GR00T 2: robot AI learns skills from human video, company says

NVIDIA on Monday, 9 November 2026 announced GR00T 2, a new open artificial intelligence model for teaching robots physical skills, at the Conference on Robot Learning (CoRL) in Austin, Texas. The company calls it a "world action model": instead of only deciding a robot's next movement, it predicts what the scene will look like in the next moments and what actions the robot should take, together. NVIDIA says the model, which has about 2 billion parameters and is built on its Cosmos family of models, learns new tasks from very few demonstrations. It also launched Isaac Lab 3.0, the latest version of its robot-training simulation software. Foxconn, Siemens and the start-up Nexuni are among the companies testing the model.

Robots in factories have long been programmed step by step for one job. Changing the job means reprogramming. The current race in robotics is to build "foundation models" for robots, large AI systems trained on huge amounts of data that can pick up new tasks with little extra teaching, the way language models answer new questions without being rebuilt.

Most recent robot models are vision-language-action (VLA) models: they look at camera images, read an instruction and output movements. NVIDIA's pitch for GR00T 2 is that predicting the future scene as well as the action helps a robot plan delicate work such as handling cloth or cables. Training on human video matters because recording robots doing tasks is slow and expensive, while video of people using their hands is plentiful. If a model can learn useful skills from people's video, the cost of teaching robots falls.

The numbers should be read carefully. The success rates come from NVIDIA's own tests, and its claim that GR00T 2 ranks among the top three models on several robotics benchmarks is also self-reported. No independent evaluation was available at launch. The partners are testing or evaluating the model; reports note that none has said it is running on a production line.

For India, where electronics assembly is growing fast and engineering teams build robotics software for global firms, open robot models lower the entry cost for start-ups and research labs that cannot afford to collect their own large robot datasets. Whether that happens depends on the licence and the release of the weights, which are still to be confirmed.

Peepals calculation: Going from one demonstration to five raised NVIDIA's reported average success from 71% to 95%, a gain of 24 percentage points, or about 34% better in relative terms (95 ÷ 71 ≈ 1.34). With no demonstrations, completing 40 of 50 tasks at least once equals 80% of tasks (40 ÷ 50 = 0.8). Our calculation from the reported figures.

Short-term: Robotics researchers and companies can begin experimenting with NVIDIA's tools and open data, while waiting for confirmed details on the model's release and licence.

Long-term: If learning from human video works as claimed, teaching robots new factory or household tasks could become much cheaper, speeding up automation in manufacturing and logistics.

Who is affected: Robot makers, factory operators, AI and robotics researchers, electronics manufacturers such as Foxconn, and workers in jobs that involve repetitive hand tasks.

What to watch next: The release date and licence of GR00T 2's model weights, and independent results on the benchmarks NVIDIA cites.

Questions and Answers