BRENNEN

NVIDIA GR00T 2

Five demonstrations per task

By NVIDIA's own test, its robot model GR00T 2 averages 95% success after five demonstrations of a task, on 50 tasks picked from those it learns that way.

Robots · Checked 13 Nov 2026 · 9 sources · 10 min

Most people can be shown a new way to fold a towel once and then do it. The showing works because it has so little to carry. Whoever is watching already knows what a towel is and how cloth behaves when it is lifted by two corners. The demonstration only has to say which fold is wanted.

A robot comes to the job with none of that. A common way to teach one a task is for a person to steer it through the motions by remote control while its cameras and joints are recorded. The field calls the steering teleoperation, and one recording of the task is a demonstration. The software that will run the robot, called a model, is then trained on the demonstrations: its internal settings are adjusted until it produces the motions itself. A model that begins with nothing has to get everything from them, down to what a towel is, and that takes many demonstrations of every task.

On 9 November 2026, at the Conference on Robot Learning, NVIDIA introduced a robot model called GR00T 2 and reported what it does with very few. Across 50 tasks, the company says, the model succeeded 71% of the time on average after one demonstration of each task, and 95% after five.

Both figures come from NVIDIA's own test, on one robot. NVIDIA chose the tasks, ran every trial, meaning every attempt, and counted the successes. A success rate depends on which tasks are picked and on what is counted as success. When one party settles both and also has something to announce, the figures may be accurate, and they stay unchecked until someone with no stake repeats the test. On 13 November 2026 the model had been introduced and was not publicly downloadable.

For five demonstrations to be enough, a model would have to arrive the way a person does, already knowing most of what the job needs. What GR00T 2 arrives having seen can be counted in hours. NVIDIA trained it on three kinds of recording. The largest is 402,730 hours of video of people at work, filmed from the worker's own point of view. A year is 365 days of 24 hours, which is 8,760 hours, so the video would take about 46 years to watch straight through. With it came 10,184 hours of demonstrations made with a gripper that a person holds in the hand, and 23,500 hours of teleoperation. This first round of training is called pretraining.

Hours of recording in pretraining, as NVIDIA gives them

Most of the hours are of people

Watched straight through, 402,730 hours is about 46 years.

NVIDIA

  1. Video of people at work
    402,730
  2. Teleoperation
    23,500
  3. Demonstrations with a hand-held gripper
    10,184

Most of those hours are of people. A video of someone folding a towel shows what the hands do and what the cloth does in reply, and it holds no commands for a robot's motors. Choosing those commands is one of two jobs GR00T 2 does. The other is to predict what its cameras are about to see. NVIDIA calls a model of this kind a world-action model, which its whitepaper, the technical report that came with the announcement, defines as a single model that "predicts both how the scene will unfold and the robot's actions."

While the robot works, the model repeats one short step. It takes in 0.8 seconds of images from up to four cameras, with the state of the robot's own body and a written instruction, and gives back a run of movements to make next and the frames it expects each camera to show. The model itself is a single network with 2 billion of the settings that training adjusts, which are called weights, and it is built on Cosmos-Predict2.5-2B, a video model pretrained for world simulation.

NVIDIA measured whether watching people helps, on Dexter, the robot behind the headline figures. It has two hands, each able to move in 22 independent ways. With the robot data held fixed, the whitepaper reports that success on 72 tasks run on the real Dexter rose from 18% to 41% as human video was added to pretraining.

Under three minutes of robot data

The headline test asks what all that watching is worth on Dexter's hands. NVIDIA assembled 50 tasks, a set it calls Dexter-50. None appears in the training data recorded on Dexter. Each is present elsewhere in the data: in video of people, in teleoperation of a different robot arm that ends in a two-jawed clamp where Dexter has a hand, or in both. Suppose folding a towel were one of the 50. The model would have watched people fold towels, and maybe a clamp, and never Dexter's own hands doing it.

NVIDIA tested the model on those tasks three ways. With the written instruction alone, the whitepaper reports success 40% of the time on average. With one demonstration of each task, made by teleoperating Dexter, it reports 71%. With five it reports 95%, and 37 of the 50 tasks succeeded on every trial. Five demonstrations amount to under three minutes of robot data, against about 46 years of video.

Success on the 50 tasks, by NVIDIA's own test

From no demonstration to five

NVIDIA whitepaper

  1. Written instruction alone
    40%
  2. One demonstration of each task
    71%
  3. Five demonstrations of each task37 of the 50 tasks succeeded on every trial
    95%

Here the likeness to a person ends. The model does not watch a demonstration and then get on with the job. A task's demonstrations are used for a further, short round of training, called fine-tuning. The 2 billion pretrained weights are frozen, a small set of extra weights is added beside them, and only the extras are trained. By NVIDIA's account one run takes about 30 minutes on a single RTX 5090, a graphics card of the kind sold for desktop computers. So "one demonstration" in the headline means one recording plus a separate fine-tuning run for that task.

Fifty tasks kept because they worked

The whitepaper says how the 50 were chosen, in a sentence where an episode is its word for a demonstration: "The 50 tasks were chosen among those the model learns with five episodes (at least 80% success), so the five-episode rate is high partly by construction." The paper does not say how many tasks were tried before 50 were kept.

Suppose a model is tried on 100 tasks, and with five demonstrations it succeeds 90% of the time on half of them and 30% of the time on the other half. Over all of them its average is 60%, the midpoint of 90 and 30. Keep only the tasks where it reached 80%, average again, and the answer is 90%. The model is the same in both sums. The second figure is true of the tasks that were kept and silent about the rest. If folding a towel had come out below the bar, the towel would not be in the test.

An average over tasks that were let in for scoring 80% or better is going to be high whatever the model does elsewhere, which is what "by construction" means. The picking does not settle how far above 80% the average lands, and the whitepaper's word is "partly": the reported 95% sits well above the bar. The 71% and the 40% were measured on the same filtered 50. The sentence quoted from the whitepaper appears on neither NVIDIA's blog nor its research page. The Next Web noted the selection in its launch report.

How far a success rate can be trusted also depends on how many trials stand behind it. A rate of 95% is one failure in every 20 attempts, and 99% is one in every 100. Say a robot does a task once a minute through an eight-hour shift, which is 480 attempts. At the first rate that comes to 24 failures in the shift, and at the second to about 5.

A short test cannot reliably tell those two robots apart. In 20 trials, the one that fails once in 20 will quite often get through without a failure. The whitepaper states that telling 95% from 99% takes about 285 trials of each model under test, and it gives no number of trials per task for Dexter-50. As Humanoid.guide reported on 10 November 2026, the robotics entrepreneur Jens Peder Kristensen distinguished research demonstrations from reliable operation across an eight-hour factory shift.

No other robot model was run on Dexter-50, so the test shows how GR00T 2 improves with demonstrations and not how it ranks. The whitepaper compares it with other models in separate experiments, which ran up to 20 trials per task. It calls that enough to resolve the large gaps between GR00T 2 and the others "but not the differences that matter in deployment".

Open, with the terms unpublished

Anyone outside NVIDIA who wanted to repeat the tests would need the model. On 13 November 2026 the research page's header read "Code Available in December", and a model search on Hugging Face, a site where models are published, showed none named GR00T 2 on its first page. The blog's verb is "introducing", and it gives no download location, license or price. NVIDIA calls GR00T 2 "an open frontier world action model built on NVIDIA Cosmos". Humanoids Daily reported that NVIDIA's Jim Fan said the weights would be published as open source in December, and that license terms were unpublished. What the word open will permit depends on those terms.

By NVIDIA's account, Foxconn is applying the model to factory automation for server manufacturing, Nexuni is testing it and Siemens is evaluating it. The Next Web reported that none of the three had said it would use the model in production, and launch coverage reports no independent test.

If the code and weights arrive in December as stated, the recipe NVIDIA describes asks for five teleoperated demonstrations, about 30 minutes of fine-tuning on a single RTX 5090 and a supported robot. The pretraining underneath used 1,024 GB300 GPUs, the class of chip a graphics card is built around, so anyone following the recipe would depend on the weights NVIDIA publishes.

An hour of teleoperation needs a robot and an operator for the hour. Video of people at work needs neither. NVIDIA has already published a 100,000-hour subset of its Human-1M collection of first-person video under the Creative Commons Attribution 4.0 license. If the rise from 18% to 41% holds outside NVIDIA's tests, video of that kind can supply part of what a robot model learns.

Repeating NVIDIA's exact test would take the weights and a Dexter. Even an exact repeat would speak for one robot, on tasks the model had already watched people or other robots do. By the paper's own statement, it is not known how little data from a target robot suffices to specify another robot body. And for a task that appears nowhere in the pretraining data, the whitepaper offers no figure.

Check it yourself

  1. Open NVIDIA's blog post and read the GR00T 2 section to confirm the verb "introducing", the 71% and 95% figures and the sentence about Foxconn.
  2. Open the GR00T 2 research page and read its header to confirm the words "Code Available in December", then find the hours of each kind of pretraining recording.
  3. Search Hugging Face for nvidia/GR00T and look for a model named GR00T 2. Whether one is listed shows whether the weights have been published.
  4. Open NVIDIA's Isaac GR00T code page on GitHub and read its front page to see whether the public code covers GR00T 2 or only earlier versions.
  5. Open the Human-1M dataset card and read the license field to confirm the 100,000-hour subset and its Creative Commons Attribution 4.0 terms.
  6. Download the whitepaper PDF, about 23 MB, and read page 8, where the 71% and 95% figures and the sentence ending "high partly by construction" sit in one paragraph.
  7. Turn to Appendix A.3 on page 41 and Figure 24 on page 42 of the same whitepaper for the definition of Dexter-50, the fine-tuning settings and the result for each task.
  8. Compare NVIDIA's pages with The Next Web's launch report to see how an outside outlet read the selection of tasks and the partner companies' statements.

Sources