Nvidia pitches a robot brain that learns from watching people. GR00T N2 trained on roughly 400,000 hours of first-person video, twenty times N1.7, and Nvidia claims it nailed 48 of 50 unseen motions after five attempts. Humanoids just levelled up.
Forget robots that need months of hand-coded choreography. On 17 November 2026, ChosunBiz reported that Nvidia unveiled the GR00T N2 humanoid robot foundation model, a “robot brain” trained on roughly 400,000 hours of egocentric video, footage shot from a first-person point of view as people go about real tasks. That is about twenty times the roughly 20,000 hours behind its predecessor, GR00T N1.7.
How the GR00T N2 humanoid robot model learns
The headline shift is architectural. According to ChosunBiz, earlier GR00T models were built on a VLA, or vision language action, structure that maps what a robot sees and hears directly to an action. N2 instead uses what Nvidia calls a World Action Model, which tries to understand how the physical world changes and predict the consequences of an action before taking it. In plain terms: less copy and paste, more cause and effect.
Nvidia’s own test numbers, as reported, are the part that will light up group chats. Across 50 new motions it had never seen, N2 succeeded at about 20 on the first try with zero prior attempts, about 40 after three attempts and about 48 after five. These are company claims, not independent benchmarks, and the full methodology was not published in the report.
| Metric | Figure |
|---|---|
| Egocentric video used in training | About 400,000 hours |
| Predecessor N1.7 training video | About 20,000 hours |
| Dataset released on Hugging Face | 100,000 hours |
| New motions succeeded, zero-shot | About 20 of 50 |
| After 3 attempts | About 40 of 50 |
| After 5 attempts | About 48 of 50 |
Open data, big partners
Nvidia also released 100,000 hours of that egocentric video as a dataset on Hugging Face, ChosunBiz said, a move that hands researchers and startups a massive training resource. The report named Boston Dynamics, ROBOTIS, Unitree and Figure among the robot makers working with Nvidia’s platform, and noted that robotics startup Skild AI recently closed a $1.4 billion Series C round. The money and the data are flowing to the same place: robots that generalise.
Why this matters for polymaths
Here is the twist. The bottleneck in physical AI is no longer just hardware; it is taste, judgment and the ability to connect robotics, vision, data rights and human behaviour. Someone has to decide which 400,000 hours matter, which tasks are safe to automate and how a humanoid should move around people. That is polymath work. The engineers who win this era will be the ones who can talk motor control on Monday, privacy law on Tuesday and product design on Wednesday.
Gen Z gets this instinctively. A generation that learned everything from first-person video is now watching robots learn the same way. The question is who gets credit for teaching them.
FAQ
What is GR00T N2?
A humanoid robot foundation model from Nvidia that, according to a 17 November 2026 ChosunBiz report, was trained on about 400,000 hours of egocentric video.
How well does GR00T N2 learn new motions?
Nvidia reported about 20 of 50 new motions succeeded with no prior attempts, about 40 after three attempts and about 48 after five. These are company figures.
What is a World Action Model?
Nvidia’s term for a model that tries to understand how the physical world changes and predict the result of an action, instead of mapping inputs directly to actions.