Nvidia GR00T 2 is a massive update: Jensen Huang says it's the 'ChatGPT moment for robotics'

Prediction is super-hot in robotics. I've seen multiple physical AI foundation models that include prediction. Now, at CoRL 2026 in Austin, Nvidia has unveiled GR00T 2, the next generation of its humanoid robot foundation model. And ... guess what ... it has prediction too.

Previous GR00T models (N1, N1.5, and successive updates) were vision-language-action (VLA) models. They processed what a robot sees, interpreted instructions, and produced movements. GR00T 2 is a World Action Model (WAM). Before it moves, it predicts how its actions will change what it sees next. A robot using GR00T2 is no longer just reacting ... it is planning.

The practical difference shows up immediately.

GR00T N1.5 handled novel tasks at 38.3%. GR00T 2 completes 40 of 50 evaluated tasks with zero task-specific training: 80% zero-shot ... incredibly impressive. Give it one demonstration per task: 71% average success. Five demonstrations: 95%.

“The ChatGPT moment for robotics is here. Breakthroughs in physical AI — models that understand the real world, reason and plan actions — are unlocking entirely new applications,” said Jensen Huang, founder and CEO of NVIDIA. “NVIDIA’s full stack of Jetson robotics processors, CUDA, Omniverse and open physical AI models empowers our global ecosystem of partners to transform industries with AI-driven robotics.”

That might just be the fastest published learning curve in humanoid robot benchmarks.

Trained on humans, not just robots

The training data shift is as significant as the architecture.

Nvidia trained GR00T 2 on 100,000 hours of human egocentric video, an open-source dataset of people doing things with their hands: picking up objects, assembling parts, manipulating everyday tools.

The logic is straightforward and aligns with companies like Figure AI: if robots need to do what humans do, training on how humans actually do it produces more generalizable behavior than training on simulated robot approximations. Nvidia's CoRL 2026 announcement positions this as a foundational data strategy shift.

The result: GR00T 2 can transfer skills learned on one robot body to a different robot body without re-demonstrating the task. If a manipulation skill is taught to an Agile Robots humanoid, GR00T 2 can apply it to a Unitree platform.

This is also huge ... and another key benefit just right out of the box.

Isaac Lab 3.0 and the simulation stack

Nvidia also launched Isaac Lab 3.0, updating its popular simulation framework.

The headline number: 2.5x the training throughput of Isaac Lab 2.3. For teams running millions of simulated robot training cycles, that's not a minor bump; it's the computational equivalent of running 2.5x more hardware at no additional cost.

Again, huge.

Isaac Lab 3.0 also integrates the Newton physics engine alongside PhysX, adding more accurate simulation of cables, cloth, and complex articulated hands. A new LEAPP tool moves trained policies from simulation to physical hardware. AlphaNIST, a system built entirely in Isaac Lab, now handles all 20 tasks on the NIST Assembly Task Board, manipulating millimeter-scale parts at over 98% training success.

Who's deploying it

The model partners are building on just got significantly more capable. And Nvidia showcased an impressive list of early adopters:

• Foxconn is applying GR00T 2 to server manufacturing, a category requiring precise, varied assembly at industrial volume.
• Siemens is evaluating it across multiple industrial applications.
• Nexuni is using it for robotic laundry handling.
• LG's CLOiD humanoid, demonstrated at Nvidia's Seoul AI Day event on the same day as the CoRL announcement, is being trained on the Isaac GR00T platform.
• Agile Robots and Unitree are both integrating with the Jetson Thor hardware platform that GR00T 2 targets.

Nvidia VP of Robotics Deepu Talla, speaking in Seoul, declared "the golden age of physical AI" and called Nvidia "a platform company" with a simple corollary: "We do not build robots."

That's a big hint: we won't compete with you if you use our physical AI platform to build robots. And it's a bit of side-eye at others who build similar software stacks, suggesting they might not choose the same limitation.

The world model framing

The shift from VLA to WAM is a big deal.

A VLA model receives input (here's what I see, here's the instruction) and outputs action. A world action model predicts what the world will look like after it acts before committing to the action. It's the difference between a chess engine that evaluates the current board and one that simulates the next five moves.

Applied to robotics, it results in a robot that doesn't just pattern-match but reasons about physical consequences of movement. That makes it faster on novel tasks and more reliable under conditions it hasn't seen before. It also enables robots that are better suited for the messy, unpredictable environments where humanoid robots need to operate: actual factories, actual homes, actual workplaces.

GR00T N1.5 launched in May 2025. GR00T 2 arrived 18 months later. In that time, zero-shot task completion jumped from 38.3% to 80%.

If that rate of improvement holds for the next iteration, the case for humanoid robots doing useful work at scale becomes much harder to argue against.

Sources

• Nvidia Newsroom — Nvidia releases new physical AI models as global partners unveil next-generation robots
• Nvidia Blog — Practice makes precise: at CoRL, Nvidia and partners reveal how robots rehearse
• The Next Web — Nvidia's GR00T 2 teaches robots delicate skills from human video
• RuntimeWire — Nvidia introduces GR00T 2 humanoid model alongside Isaac Lab 3.0
• Techgenyz — Nvidia AI Day Seoul 2026: Samsung, LG and Hyundai expand AI work