Teaching a robot a new job used to mean weeks of programming, or months of collecting robot training data. NVIDIA's GR00T 2, introduced on 9 November at the Conference on Robot Learning in Austin, claims to cut that down to an afternoon. Show it a task five times, fine-tune for about 30 minutes on a single RTX 5090 graphics card, and NVIDIA reports 95 percent average success across 50 dexterous tasks. The trick is that the robot learned most of what it needs before it ever saw your task, by watching people. I have read the announcement and the research page carefully, because the headline number is real, and so is the asterisk next to it.
How it learns from watching people
GR00T 2 is what NVIDIA calls a world action model, built on its Cosmos world-model platform. It watches recent camera video, reads the robot's current state and a language instruction, and predicts what the scene will look like next and which actions cause that change. The pretraining mix is the story:
- 402,730 hours of first-person video of people at work, filmed from the worker's own point of view.
- 10,184 hours of demonstrations made with a handheld gripper across 2,897 tasks.
- 23,500 hours of robot teleoperation data.
On NVIDIA's own 50-task test set, called Dexter-50, run on a two-handed robot with 22-degree-of-freedom hands, the reported results were 40 percent average success with a written instruction alone, 71 percent after one demonstration of each task, and 95 percent after five demonstrations. In one striking test, a human alone showed a robot how to line a laboratory funnel with filter paper and wet it with a wash bottle, and the five-fingered robot completed it in 65 percent of trials, having never been teleoperated on that job. A smaller variant, GR00T 2 Turbo, answers in 124 milliseconds on NVIDIA's Jetson Thor robot computer, fast enough for real-time control on the robot itself. NVIDIA has also opened 100,000 hours of that human video as the Human-1M-100K dataset under a CC BY 4.0 licence, and Foxconn is already applying the model to server manufacturing, with Siemens evaluating it for industrial use.
The angle nobody is spelling out: automation gets affordable when teaching gets cheap
The cost of a robot was never just the machine. It was the engineer who programmed it, the weeks of setup every time the job changed, and the fact that only big factories could afford any of that. If five demonstrations and a 30-minute fine-tune genuinely teach a new task, the maths flips: a small workshop, a packaging line, a repair bench could retrain a robot the way you train a new worker, by showing it the job a few times. That is the promise that matters here, and it is bigger than any single benchmark. I wrote this week about Genesis AI raising $500 million to teach robots to use their hands like humans, and about 100 humanoid robots heading for Aptiv's factories, and GR00T 2 is the same race from a different angle: the winner will be whoever makes the teaching cheapest, not whoever builds the fanciest hand.
The asterisk, stated plainly
Two honest caveats. First, the whitepaper itself says the 50 tasks were chosen among those the model learns with five demonstrations at a success rate of at least 80 percent. The 95 percent average is real, and it is high partly by construction: the test set was filtered for tasks the model could already learn. NVIDIA does not say how large the pool was, so nobody outside the company knows how selective that filter was. Second, every figure here is NVIDIA's own, on NVIDIA's own test, with no independent replication yet. And the model itself is not downloadable today: the research page lists code availability for December 2026, and no date has been given for the model weights. What is genuinely open right now is the dataset and the published results, not the model.
Bottom line
GR00T 2 is the clearest sign yet that robot teaching is moving from code to demonstration, and that shift, if it holds up outside NVIDIA's lab, changes who can afford a useful robot. The 95 percent figure deserves its asterisk, and the model deserves its December test, when outside researchers finally get the code. Until then, watch what Foxconn and Siemens actually ship with it, because a factory floor is the one benchmark nobody can filter.
Sources: NVIDIA, Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation, 9 November 2026, blogs.nvidia.com; NVIDIA Research, GR00T 2 project page, research.nvidia.com.
Originally published on Ai Game Changer Pro.