NVIDIA's robot learned its job by watching videos, and the 95% number has a catch
Hello guys,
A robot that learns a new job by watching videos of humans doing it. That is what NVIDIA showed on Monday, and the headline number is hard to ignore: 95% success after five demos.
But the 95% needs a second look, and NVIDIA's own paper says why. So here is my take: this is the most interesting robot news of the year, and the number everyone is quoting is the least honest part of it.
GR00T 2 is the name, a robot brain built on NVIDIA's Cosmos world models, shown at the Conference on Robot Learning in Austin. It watches a short clip, reads the instruction, checks the robot's hands, then predicts the next few seconds of video and the motor commands to match.
The training data is the real story for me. NVIDIA fed it 402,730 hours of first-person human video. That is people doing ordinary things with their own hands. Add 10,184 hours of gripper demos and 23,500 hours of robot teleoperation, and you get a model that starts with a rough idea of how hands work before it ever touches a robot arm.
Then the numbers. On 50 tasks with no training data from that robot, GR00T 2 got 40% right with no demo at all. Show it one demo of each task and it jumps to 71%. Show it five demos and it hits 95%.
That ladder is the whole point, I think. One demo nearly doubles the score. Five demos nearly fix it. The future of teaching robots is showing them, the way you would show a new worker on their first day.
Now the catch. And it is a big one, because it comes from NVIDIA's own whitepaper.
Those 50 tasks were picked from a bigger pool. The paper says it plainly: the tasks were chosen among the ones the model can learn with five episodes at 80% success or better. So the 95% is real, but it is measured on tasks that were already known to be learnable. NVIDIA does not say how big the pool was. The demo reel is a highlight reel.
Then there is the word open. NVIDIA calls GR00T 2 an open frontier model, and part of that is true. They did release 100,000 hours of the human video as Human-1M-100K, free to use under a Creative Commons licence, even for commercial work.
But the model itself is not out yet. The research page says the code arrives in December. No weights link, no licence, no price, no date. Open with a waiting room is not quite open.
The part that got me most is not the 95% at all. It is what happens when you pair GR00T 2 with a planner that thinks. NVIDIA ran toy assembly tests where a reasoning model watches, plans, and corrects the robot as it goes. With the planner, using GPT-6 Astra, the robot finished 97 out of 100 runs. Without it, 23 out of 100.
That tells me the robot hand was never the hard part. The hard part was knowing what to do next, and when to stop and try again. A robot that can watch and copy is a big step. A robot that can notice its own mistake and fix it is the real jump.
And here is the honest bit that keeps me grounded. NVIDIA also showed GR00T 2 Turbo running on their Jetson Thor chip, doing simple jobs like scooping popcorn and clearing a table. It works. But wiping a towel underfoot still fails about half the time. Half. A job any of us does without thinking, and the robot gets it wrong every second try.
Foxconn is testing it for server manufacturing, Siemens for industrial work, Nexuni for laundry. A small team called MicroAGI used it to swap Ethernet cables. All testing, none of it a real production line yet. A factory does not need a robot that is right 95% of the time. A factory needs a robot that is right 99.9% of the time, on a bad day, at 3 in the morning, when nobody is watching. One dropped part in twenty still stops a line.
Isaac Lab 3.0 came out with it, the simulator where robots practice. New physics for cables and cloth, agents that set up their own training scenes, up to 2.5 times faster training than the last version. Boring on paper, huge in practice.
So where do I land? Watching is how doing starts. The 95% is a picked number and the model is not even out yet, and I still think this is the most important robot release of the year. The reason is the idea underneath it: show a robot, do not program it. That is how humans learn a job, and it might be how robots finally learn one too.
The robot that wins will be the one that never drops the towel.