How did you learn to fold a towel? Probably by watching someone else do it a few times. Nobody grabbed your wrists and moved your hands for you, over and over.
Robots, until recently, mostly learned the hard way. People had to steer them by remote control to record each example. On November 9, NVIDIA showed a different route. Its new model, NVIDIA GR00T 2, picks up a big chunk of its skills by watching videos of humans doing everyday things.
It’s one of the biggest physical AI stories of the week. And honestly, it’s a lot more interesting than another humanoid doing a dance video.
So what did NVIDIA GR00T 2 actually show?
NVIDIA unveiled GR00T 2 at CoRL, a major robot-learning conference in Austin, Texas. The company calls it a “world action model.” That’s a fancy way of saying it predicts how its own moves will change what it sees, before it moves.
Think of it like this. Before you reach for a cup, you already picture your hand closing around it. GR00T 2 tries something similar. It forecasts a short clip of future video and the matching robot actions at the same time.
The model also works across different robot bodies. NVIDIA’s research page lists simple two-finger grippers, dexterous robot hands and humanoids. There’s a smaller version, too, called GR00T 2 Turbo, which runs on the robot itself.
Why learning from human video is a big deal
Here’s the core problem in robotics. Good robot data is slow to collect. Someone has to operate the robot by hand for every example. We looked at one version of this in our piece on Korea’s humanoid data factory, where people teach robots by remote control.
Human video is a different story, since you don’t need a robot in the room to record it. So NVIDIA trained GR00T 2 mostly on people. The mix includes 402,730 hours of first-person human video, 10,184 hours of handheld gripper demos and 23,500 hours of robot remote-control data.
Then comes the clever part. NVIDIA says the model can carry a skill from a person, or from one robot, over to a robot with a different body. In one test, a robot hand learned to line a lab funnel with filter paper from human demos alone. It succeeded in 65% of trials, without a single remote-controlled example.
NVIDIA is also giving away 100,000 hours of that human video as an open dataset called Human-1M-100K. That’s a gift to every robotics startup that can’t afford to film its own.
The numbers, and the fine print
The headline results look great. NVIDIA tested 50 tasks the target robot had never trained on. Without any practice, GR00T 2 succeeded at least once on 40 of them. With one demo per task, average success hit 71%. With five demos, it reached 95%.
Those five demos add up to less than three minutes of robot data per task. Then a 30-minute fine-tune on a single RTX 5090 graphics card finishes the job. That’s one graphics card, not a giant server farm.
Now the fine print. The Next Web points out that NVIDIA’s own paper says those 50 tasks came from ones the model could already learn well. So the 95% is an average on friendly ground, not a promise for any random job. These are also NVIDIA’s own tests, and we haven’t seen independent checks yet.
Who’s already trying NVIDIA GR00T 2?
A few big names are kicking the tires. According to The Next Web, Foxconn is using it to help build AI servers. Siemens is testing it for factory work, while a company called Nexuni is trying it on laundry.
None of them has said it’ll use the model in production yet, though. That’s normal at this stage, so don’t read too much into it.
Some of the demos are fun, too. As Interesting Engineering reports, the smaller Turbo model ran on a Unitree G1 humanoid and served popcorn in all 48 trials. It also cleared tables in 93% of trials. In another test, robot hands seated GB300 memory modules in 85% of tries. Honestly, that last one impressed us most, because it’s fiddly, precise work.
UC Berkeley professor Ken Goldberg, who worked with NVIDIA on two of the projects, went even further: “To me, agentic robotics is the most exciting technological development since the internet in 1995.”
What it means for factories, and for us
So why should you care if you don’t run a robot lab? Because cheaper training is what turns robots from demos into products. If a factory can teach a robot a new task with five demos and half an hour of computing, smaller manufacturers suddenly get a seat at the table.
That’s the part we find exciting for places like Armenia. Big automation projects have mostly been a rich-company game. Open models and open datasets lower the entry ticket for local startups and engineering teams. We saw investors chase the same idea with Physical Intelligence, which is betting big on general robot brains.
NVIDIA also released Isaac Lab 3.0, its open-source tool for training robots in simulation, alongside the model. The code for GR00T 2 should follow in December. In other words, builders won’t have to wait long to try it themselves.
So, will robots learn your job by watching you?
Not tomorrow. NVIDIA GR00T 2 is a research model, and the strongest numbers come with caveats. But the direction is clear. Robots are starting to learn the way we do: by watching first, then practicing a little.
That raises a question we can’t stop thinking about. If every video of a person doing a task becomes robot training material, who gets paid for that footage? One startup, Maxinsights, already pays people to film household chores. Expect that market to grow.
So here’s our question for you. Would you film yourself doing your daily work if it helped train the next generation of robots, and what would you want in return?