Models
Nvidia GR00T 2 reaches 95% average success on 50 robot tasks after five demonstrations each and a 30-minute fine-tune; the figures are Nvidia’s own, on tasks picked because the model could learn them, and human video is most of the training mix
Nvidia says GR00T 2 hits 95% average success on 50 tasks after five demos each. The tasks were chosen as learnable, the figures are Nvidia’s own.
Nvidia unveiled GR00T 2 on Monday, November 9 at the Conference on Robot Learning in Austin, an open robot model that learns partly from first-person human video. The headline is 95% average success after five demonstrations per task. That number is real but narrow: it is Nvidia’s own measurement, on 50 tasks that were picked because the model can learn them, and the demonstrations are of the task on the target robot, not just a person on video.
The essentials
- What it is: Nvidia calls GR00T 2 an open “world action model”, built on its Cosmos world models, which predicts how an action will change what the robot sees before it moves. Foxconn, Siemens and Nexuni are reported to be testing it, per The Next Web.
- The numbers: on 50 tasks unseen on the target robot, it completed 40 of 50 at least once with no task-specific demonstrations, which The Neuron stresses is coverage, not an 80% success rate. Zero-shot trial success was 40%, 71% with one demonstration and 95% with five, after about 30 minutes of fine-tuning on a single RTX 5090 and under three minutes of robot data per task.
- The caveat: per The Next Web, Nvidia’s paper notes the tasks were picked from ones the model could already learn well, and Pebblous reads the selection as tasks learnable at 80% or better from five episodes. Treat 95% as a ceiling for favorable tasks, not a field average.
- The data: the training mix is 402,730 hours of first-person human video, 10,184 hours of handheld-gripper demonstrations and 23,500 hours of robot teleoperation, so human video is roughly 92% by Pebblous’s and FourWeekMBA’s arithmetic. Nvidia released 100,000 hours as Human-1M-100K, reported under CC BY 4.0. Adding human video raised real-robot task progress from 50% to 70% across 72 tasks, per Superpower Daily.
Learning from video is only part of the recipe
The human video supplies most of the hours, but robot teleoperation and handheld-gripper data remain in the mix, and each new task still needs a handful of demonstrations on the target robot. That is far less than a long training campaign, but it is not learning from watching alone.
What would change the picture
Independent evaluations on tasks Nvidia did not choose, and reliability over long runs. Completing a task once is a long way from running all day on a factory line. This is a reading of reported figures, not a hardware evaluation.
What is not established
- Independent replication: none seen in the sources.
- Performance on tasks not selected for learnability: not reported in what we read.
- Long-run reliability on a production line: not shown.
- Nvidia’s paper and research page: we did not read them directly.
Summary table
| Point | What is stated as of November 7, 2026 | Source |
|---|---|---|
| Tasks completed at least once | 40 of 50 with no task demonstrations | The Neuron |
| With one demonstration | 71% average success | The Next Web |
| With five demonstrations | 95% average success, 30-minute fine-tune on one RTX 5090 | Interesting Engineering |
| Training mix | 402,730 h human video, 10,184 h gripper, 23,500 h teleoperation | Pebblous |
| Effect of human video | Real-robot progress 50% to 70% across 72 tasks | Superpower Daily |
FAQ
Does GR00T 2 learn only by watching people?
No. Human video is most of the pretraining hours, but it also uses gripper and teleoperation data and needs a few demonstrations per task.
Is 95% its success rate on any task?
No. It is an average over 50 tasks that Nvidia selected as learnable, measured by Nvidia.
Can I use the video data?
Nvidia released 100,000 hours as Human-1M-100K, reported under CC BY 4.0.
Method and limits
Sources checked on November 10, 2026: The Neuron, Interesting Engineering, Pebblous, The Next Web and Superpower Daily, read as search extracts. Nvidia’s paper and research page were not read directly. Success figures are Nvidia’s, not independently validated.
Sources
- The Neuron: NVIDIA GR00T 2: Human Video, Robot Learning and Caveats (verified 2026-11-08)
- Interesting Engineering: NVIDIA's GR00T 2 lets robots learn tasks from humans, not machines (verified 2026-11-08)
- Pebblous: GR00T 2 Training Data: Only 5.4% Comes From Robots (verified 2026-11-08)
- The Next Web: Nvidia's GR00T 2 teaches robots delicate skills from human video (verified 2026-11-08)
- Superpower Daily: Nvidia introduces GR00T 2, a robot AI trained on human video (verified 2026-11-08)
Spotted an error in this article? Report it · Our corrections policy