NVIDIA GR00T 2: a robot model trained on human video, with code promised for December

NVIDIA's GR00T 2 learns hand skills from human video, and the company says it can finish 95% of tasks after seeing five demonstrations of each. The robot foundation model was unveiled at the Conference on Robot Learning in Austin on Monday, November 9, 2026, with a whitepaper and a promise of code in December. The claims are specific and testable, which is the good news. The bad news for now is that nobody outside NVIDIA and its partners can test them.

What GR00T 2 is

NVIDIA's GEAR lab calls GR00T 2 a "world action model". The idea is that the model predicts future video and robot actions together, so it learns how a physical interaction unfolds, not only which motor command usually follows a picture. The NVIDIA research page says it runs on parallel grippers, dexterous hands and humanoids. A smaller variant, GR00T 2 Turbo, has about 2 billion parameters and is built to run on the robot itself, specifically the Unitree G1 humanoid.

This is the successor to GR00T N1, which NVIDIA described in a March 2025 paper as an open foundation model for humanoid robots with a dual-system design: a vision-language module to interpret the scene and a diffusion transformer to generate actions in real time. GR00T 2 keeps the open posture but changes the training recipe, adding a large pool of human video.

That pool is the story. NVIDIA says it trained on robot data and on human data, including 100,000 hours of first-person (egocentric) footage that it is releasing publicly as the Human-1M-100K dataset on Hugging Face. People are much cheaper to film than robots are to operate, so the bet is that skills can be learned from video of human hands and then transferred to machines with different bodies.

The results NVIDIA reports

The headline figures come from NVIDIA's own research page and from The Robot Report's account of the CoRL announcement. With one demonstration per task, GR00T 2 averaged 71% success across 50 tasks, rising to 95% with five demonstrations. NVIDIA says five demonstrations amount to under three minutes of robot data, followed by a 30-minute LoRA fine-tune on a single RTX 5090 graphics card.

On tasks it had never been shown, NVIDIA's page reports 40% success across 50 tasks, while The Robot Report says the model completed 40 of 50 tasks at least once without task-specific demonstrations. Those are different measures: one is a rate, the other a count of tasks with at least one success. We flag the difference because a model that succeeds once is not a model that succeeds reliably.

Other results on the page include 85% on seating memory modules and 65% on toy truck assembly using two 22-degree-of-freedom Sharpa hands; 65% on a filter-paper task learned only from human video; and 92% instruction-following on a YAM arm when visual steering was added, up from 40% with the written instruction alone. The Turbo variant managed 124 milliseconds per prediction on a Jetson Thor computer and cleared a table 93% of the time on the G1.

The caveats in the fine print

The Next Web, reporting on the announcement, passes along a caveat from NVIDIA's own paper: the evaluation tasks were chosen from ones the model could already learn well. That is a normal research choice and an important one. It means 95% describes the model on tasks suited to it, not on the full mess of a factory or a kitchen.

The deployments are also early. Foxconn is applying GR00T 2 to server manufacturing automation, Siemens is evaluating it for industrial use, and Nexuni is testing it for robotic laundry. The Next Web notes that none of the three has said it will use the model in production. NVIDIA's research page says code is "Available in December", so as of today there is nothing to download. The page does not state a license for the code, the weights or the dataset, and it does not say whether model weights will be released at all.

So the word "open" needs a footnote. The dataset is public. The model, for now, is a whitepaper.

Isaac Lab 3.0 and the benchmarks

NVIDIA announced companions. Isaac Lab 3.0 is the latest version of its open-source robot learning framework, and NVIDIA reports up to 2.5 times the training throughput of version 2.3 on a vision-based dexterous manipulation benchmark when using the Newton physics solver. Isaac Lab-Arena 0.4 is a developer preview for scaled policy evaluation. Two new benchmarks, DexBench with RLWRLD and VABAR from UC Berkeley and NVIDIA, are meant to give the field shared tests.

Shared tests matter more than any single score. A robotics field where every lab grades itself on its own tasks cannot tell progress from marketing. NVIDIA also says GR00T 2 ranks among the top three models in its categories on the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks, which is a claim others can check once the model ships.

An outside voice

Ken Goldberg, a UC Berkeley robotics professor who worked with NVIDIA on two of the projects, told The Next Web: "To me, agentic robotics is the most exciting technological development since the internet in 1995,". He has a stake in the work, so read it as enthusiasm from a collaborator, not an independent audit.

How we read it

The honest summary is that NVIDIA has published a credible, well-documented step toward robots that learn from video, with unusually concrete numbers, and has not yet let anyone else touch it. Compare that with how the same week treated Google DeepMind's robot model, which is limited to partners and reports its own pilots. In both cases the vendor holds the evidence.

What we would do: wait for the December code drop, then look for three things. First, third-party results on tasks the authors did not pick. Second, the license, which decides whether the word open means anything for businesses. Third, a named production deployment with a failure rate attached. Until then, treat 95% as a measured result on a friendly test, and treat the human-video approach as the genuinely interesting idea.