Index of Plates

Plate VI

NVIDIA's Coming Robots, and How They Change the World

Physical AI

22 OCTOBER 20268 Min Read

NVIDIA doesn't sell a robot you can buy. It's building the brains, simulators and safety systems that almost every new robot runs on. Here's what it announced in 2026, and what that means for the next decade.

The strange thing about NVIDIA's robots is that NVIDIA doesn't really make robots.

It makes the parts that matter most: the computer in the robot's head, the model that decides how it moves, the simulator it learns in, and now the safety system that stops it from hurting anyone. Unitree, Agile Robots, Figure, Agility, Boston Dynamics and dozens of others build the bodies. A growing share of them run on NVIDIA underneath.

That's the same playbook NVIDIA ran with AI in the data center. Stay out of the end product, make yourself the thing every product depends on. In 2026 it ran that playbook on robotics, across three big events: GTC in March, GTC Taipei in June and GTC Berlin this week.

At the March keynote Jensen Huang put it bluntly: "Physical AI has arrived — every industrial company will become a robotics company." Here's what's actually behind that line.

The Stack, in Plain Terms

It helps to see the whole thing before looking at the new pieces. NVIDIA's robotics platform has five layers:

  • Jetson Thor, the onboard computer. A Blackwell GPU with 128 GB of memory and 2,070 FP4 teraflops, small enough to sit inside a humanoid.
  • Isaac GR00T, the robot foundation models. You give them camera images and an instruction, and they output motor commands.
  • Cosmos, world models that generate and reason about physical scenes, mostly to produce training data.
  • Isaac Sim, Isaac Lab and Newton, where robots are simulated and trained before they touch real hardware.
  • Halos, the newest layer: a safety platform for robots and vehicles.

Each one is useful by itself. Together they're the closest thing robotics has to a standard toolchain.

GR00T: The Brain Is Getting Better Fast

GR00T N1.7 went into early access in March with a commercial licence. It's a 3-billion-parameter vision-language-action model built in two halves: a 2-billion-parameter reasoning model that works out what to do, and a diffusion transformer that turns that plan into joint movements. It's built for detailed hand work, down to individual fingers on 22-joint hands. That's what small-parts assembly needs, and what earlier robot models were worst at.

The interesting part is the training data. N1.7 was pre-trained on more than 20,000 hours of video of people using their hands, filmed from their own point of view, instead of the few thousand hours of robot teleoperation its predecessor used. NVIDIA reports that going from 1,000 to 20,000 hours of human video more than doubled average task completion.

That matters more than any benchmark. The biggest problem in robotics is that there's no internet-scale dataset of robots doing things. If robots can learn dexterity from watching people, that problem starts to shrink.

GR00T N2, previewed in March and due by the end of this year, changes the architecture again. NVIDIA calls it a world action model: rather than mapping images straight to actions, it also learns to predict how the world will respond to them. NVIDIA claims it succeeds at new tasks in unfamiliar environments more than twice as often as leading vision-language-action models. That's NVIDIA's own number and it should be read that way, but the direction is right. A robot that can predict consequences generalises better than one that only copies.

Cosmos 3: Training Data You Don't Have to Film

Launched in June at GTC Taipei, Cosmos 3 is an open world foundation model that combines three jobs in one architecture: understanding scenes, generating them and predicting actions. It comes in sizes from a large Super model down to an Edge version meant to run on the robot itself.

For most teams its real job is synthetic data. Filming a robot failing to grab a slippery bottle ten thousand times is slow and expensive. Generating ten thousand realistic variations of that scene, with different lighting, clutter and object positions, is a compute bill. ANYbotics, for example, has generated more than 1,500 synthetic video clips to train its industrial inspection robots.

A Humanoid Anyone Can Build On

The most concrete announcement this year is the Isaac GR00T Reference Humanoid, NVIDIA's first open humanoid reference design, announced in June:

  • A Unitree H2 Plus body, about six feet tall
  • Two Sharpa Wave tactile hands, for 75 degrees of freedom across the whole robot
  • A Jetson AGX Thor computer onboard
  • Roughly three hours of battery life, and a 7 kg rated arm payload

It's sold through Unitree from late 2026, aimed at research labs. Stanford, ETH Zurich, Ai2 and UC San Diego are among the first users.

This week in Berlin, Agile Robots added a second reference platform, the Agile ONE SN. It's smaller, 1.4 m and 40 kg, with 20-joint hands that sense touch and a planned integration with Halos. It's due in early 2027.

Why does a reference robot matter? Because until now every humanoid lab built on a different body, so nobody's results transferred. A shared, open platform means a policy trained at ETH can run at Stanford. That's how a research field speeds up: same hardware, shared data, comparable results.

Berlin: From Research to Factories

The October announcements were less flashy and more important. They're about making robots shippable.

Isaac Lab 3.0 is now generally available. It's the open-source framework for training robot policies, from a desktop up to a data center, on the Newton and PhysX physics engines.

Isaac OS is a new open-source runtime designed to behave identically in simulation and on the real robot, in real time. Anyone who has watched a policy that worked in simulation fall apart on hardware knows why that matters. NVIDIA hasn't given a release date yet.

Halos is the piece I'd watch most closely. It's a full safety platform for robots, built on NVIDIA's IGX Thor hardware. Agility, Agile Robots, KION and others are integrating it, and certification bodies including TÜV SÜD are working on how to certify industrial physical AI. Safety certification is the gate between "pilot" and "deployed at scale". Nobody puts humanoids next to workers without it.

There's also a lot of supporting groundwork: an open library of more than 2,000 simulation-ready robot and factory assets from FANUC, Universal Robots, Franka and others, and a new human-activity dataset with synchronised video and 3D hand tracking.

Who's Already Using It

This isn't just a roadmap. Some of it is in use today:

  • Industrial arms. FANUC, ABB, Yaskawa and KUKA have more than 2 million robots installed between them. All four are adding NVIDIA simulation and Jetson compute, so factories can test a production line virtually before installing it.
  • Warehouses. KION and Accenture are training autonomous forklifts for GXO in digital twins. Toyota Material Handling Europe reports cutting deployment time by up to 50% with Isaac Sim.
  • Hospitals. CMR Surgical, Johnson & Johnson MedTech and Medtronic are using NVIDIA simulation and compute to train and validate surgical robots before they reach patients.
  • Hotels. Bellboy's delivery robot Billie runs on Jetson and GR00T and is already working in hotels run by Marriott, Hilton and IHG.
  • Entertainment. Disney trained its walking Olaf robot for Disneyland Paris with a physics simulator built into Newton.

How This Changes the World

Here's my honest read.

1. Building a robot gets much cheaper. Today most of a robotics startup's money goes into rebuilding the same foundations: perception, simulation, a control stack, data pipelines. When those come as open models and open frameworks, a small team can spend its effort on the one task it's actually solving. That means far more robots aimed at far more specific jobs.

2. Simulation becomes the factory before the factory. Production lines, warehouses and operating theatres will be built and tested as digital twins first. Mistakes get made in software, where they're cheap.

3. Robots learn from us. If dexterity scales with human video, every person filmed doing skilled work with their hands becomes potential training data. That's powerful, and it raises real questions about consent, ownership and whose skills get automated first.

4. Safety becomes the real bottleneck. The models are improving faster than our ability to certify them. Whoever solves certification for learned systems, NVIDIA, regulators or someone else, decides how fast robots move out of fenced-off cells and into shared spaces.

5. Robotics standardises around NVIDIA. This is the part worth being clear-eyed about. An open stack is still a stack, and nearly all of it is tuned for NVIDIA hardware. That's great for developers right now. It also means one company sits underneath much of the next generation of machines, much as it does for AI today.

The Gap That's Left

None of this means humanoids will be in your home next year. The reference humanoid runs for about three hours on a charge. GR00T N2 isn't released yet. Isaac OS has no ship date, and Halos integrations are still "planned" for some partners. NVIDIA's own releases carry the usual note that products are available when and if available.

What's changed is the shape of the problem. Five years ago every robotics team started from zero. In 2026 a team can pick up a reference body, an open brain, a world model for data and a simulator, and start on the part that's actually theirs.

That doesn't change the world overnight. It changes how fast everything after it happens.

Sources

Written & engraved by Mayank Bairagi