Nvidia Expands Its Physical-AI Stack With New Isaac GR00T and Cosmos Models for Robots

Nvidia Expands Its Physical-AI Stack With New Isaac GR00T and Cosmos Models for Robots

Nvidia is pressing its bet that the next great AI platform will not live on a screen but inside a machine that can move. Across a run of announcements anchored at its GTC conference in March and amplified again during National Robotics Week in April, the company rolled out new open Isaac GR00T foundation models that let robots parse natural-language instructions and chain together multi-step tasks, alongside new Cosmos world models that manufacture the synthetic data needed to train those robots at scale. The through-line is what Nvidia now calls "physical AI" — a full-stack effort spanning humanoids, manipulation and simulation.

The pitch is unusually blunt for a chipmaker. "Physical AI has arrived — every industrial company will become a robotics company," said Jensen Huang, Nvidia's founder and CEO. "Nvidia's full-stack platform — spanning computing, open models and software frameworks — is the foundation for the robotics industry, uniting a worldwide ecosystem to build the intelligent machines that will power the next generation of factories, logistics, transportation and infrastructure."

What Nvidia Actually Shipped

At the center is Isaac GR00T, Nvidia's open family of vision-language-action (VLA) models — systems that map camera input and spoken or written instructions directly to continuous robot movements. The latest release, GR00T N1.7, is a 3-billion-parameter open reasoning VLA model now available in early access with commercial licensing. Nvidia says it brings generalized robot skills, including advanced dexterous control, to production-ready deployments. An earlier N1.6 release integrated Cosmos Reason, a customizable reasoning model that acts as the robot's "deep-thinking brain," turning vague instructions into step-by-step plans using prior knowledge, common sense and physics.

The second pillar is Cosmos, Nvidia's line of world foundation models. At GTC the company announced Cosmos 3, which it described as the first world foundation model to unify synthetic world generation, vision reasoning and action simulation in a single system. The point is to generate diverse, physically plausible training data — from text, image and video prompts — so robots can learn behaviors they would otherwise have to acquire slowly and expensively in the real world. Nvidia's open Cosmos world models have been downloaded more than 3 million times, and its companion Physical AI Dataset on Hugging Face has been downloaded over 4.8 million times.

Rounding out the stack are simulation tools: the general availability of the open-source Newton physics engine 1.0, co-developed with Google DeepMind and Disney Research; Isaac Sim 6.0; and Isaac Lab 3.0 for large-scale robot learning. Rev Lebaredian, Nvidia's vice president of Omniverse and simulation technology, framed the trio as a division of labor. "Humanoids are the next frontier of physical AI, requiring the ability to reason, adapt and act safely in an unpredictable world," he said. "Developers now have the three computers to bring robots from research into everyday life — with Isaac GR00T serving as robot's brains, Newton simulating their body and Nvidia Omniverse as their training ground."

The Ecosystem Nvidia Is Recruiting

The models arrive with a long list of adopters that reads like a census of the robotics industry. FANUC, ABB Robotics, YASKAWA and KUKA — which together have a global installed base exceeding 2 million robots — are integrating Nvidia's Omniverse libraries and Isaac simulation frameworks into their commissioning tools. Humanoid pioneers including 1X, Agility, Boston Dynamics, Figure and NEURA Robotics are building on Cosmos, Isaac Sim and Isaac Lab. AGIBOT, LG Electronics and others are adopting GR00T N models for industrial humanoid deployments. Nvidia has also partnered with Hugging Face to fold Isaac and GR00T into the open-source LeRobot framework.

Huang used his keynote to preview what comes next: GR00T N2, a next-generation robot foundation model built on a new "world action model" architecture derived from Nvidia's DreamZero research. Nvidia claims the model helps robots succeed at new tasks in unfamiliar environments more than twice as often as leading VLA models, and says it currently tops the MolmoSpaces and RoboArena leaderboards for generalist robot policies. It is slated to arrive by the end of the year.

Why It Matters

The robotics field is undergoing the same shift that reshaped language AI: from bespoke, task-specific systems to general-purpose foundation models that can be adapted with far less data. A single VLA model that understands "put the onion in the wooden bowl" and executes it — rather than hand-coded routines for each task — is the robotics equivalent of a large language model, and it changes who can build a capable robot.

Synthetic data and the sim-to-real transfer are the enabling trick. Real-world robot data is slow, costly and inconsistent to gather; every warehouse, farm and factory floor differs. World models like Cosmos let developers generate near-limitless training scenarios and validate policies in simulation before a robot ever touches hardware. Nvidia and its partners argue that robots trained on physics-aware world models need dramatically less real-world data to perform reliably in conditions they have never seen — the difference between a demo and a deployable product.

Strategically, this is Nvidia running the same playbook that made it dominant in generative AI: sell the compute, but also own the models, simulation frameworks and developer ecosystem so the whole industry standardizes on its stack. By keeping GR00T and Cosmos open and free while monetizing the chips underneath — from DGX-class training systems to Jetson Thor modules on the robot itself — Nvidia positions itself to benefit whether or not any single robotics company wins.

What to Watch

The open questions are about proof, not ambition. Whether GR00T N2 lands on schedule and delivers its claimed generalization gains outside benchmark conditions will signal how fast VLA models are maturing. Watch, too, for real production deployments — Foxconn's assembly lines, KION's autonomous forklifts, humanoid pilots at Figure and 1X — to see whether simulation-trained policies hold up on the factory floor. Nvidia has built the platform and recruited the ecosystem; the next chapter is measured in robots that actually ship.