In early 2024, Jensen Huang walked onto the GDC stage with two robots made by Disney that had learned to walk using NVIDIA’s AI. Most people in the audience probably thought it was a “interesting demo.” It didn’t quite hit like the announcement it was. Because Huang wasn’t just showing something new. A platform called GR00T was the first form of something bigger. It was NVIDIA’s quiet but serious bet that it could become the standard intelligence layer for a whole generation of physical machines.
Generalist Robot 00 Technology is what GR00T stands for. The name is a bit silly, but the goal behind it is serious. It’s a simple idea, but it’s really hard to put into practice: create a single AI model that can be the thinking brain for humanoid robots with different body types, environments, and tasks, without having to program each movement from scratch. People who study this have been looking for it for years. The GEAR lab at NVIDIA has called it a “moonshot mission” to achieve embodied AI, which is the bigger goal of letting AI systems live and act in the real world instead of just processing data on a server.
Vision-Language-Action, or VLA, is what GR00T is all about. That means it processes both pictures from the camera and instructions given in natural language, then turns what it has learned into motor commands that make the body move smoothly and in sync. It’s like a robot reading the room. If it hears “put the cup on the left side,” it sees a table and figures out how to move its arm to make that happen. The newest version, GR00T N1.6, also uses a layer of reasoning called Cosmos Reason. This lets the system understand what’s going on and decide what to do before it does anything. That is a difference that matters.
The platform is interesting not only because of the model, but also because of what NVIDIA has built around it. Isaac Sim lets you test how robots behave in virtual worlds before they are used on a real factory floor. You can train at Isaac Lab. Isaac Teleop records how people move so that robots can learn from it.
The Jetson Thor chip, which has a Blackwell GPU and 128 gigabytes of unified memory, runs it all on the device itself. It was designed for real-time inference on a robot’s body instead of a faraway data center. In robotics, it’s more important than in almost any other field that hardware and software are made to work together.
That’s when NVIDIA announced a reference humanoid robot in May 2026. It was made up of a Unitree H2 Plus chassis and five-fingered Sharpa tactile hands that were running GR00T software on a Jetson Thor. It’s almost six feet tall, weighs 150 pounds, and its hands and body can move in 75 different ways. The design is already being used for research at Stanford, ETH Zurich, and UC San Diego.
In tech, the goal of a reference design has always been the same: make it easier for people to get started, build a common base, and let a larger ecosystem build on top of it.
The comparison to Android isn’t a mistake. NVIDIA has made it clear that it wants to become the platform that manufacturers build on for robots, similar to how Android did for smartphones. Instead of competing to make robots itself, NVIDIA wants to become the platform that manufacturers build on. That point of view makes sense in some ways.
NVIDIA already gives you the computer. The tools for simulation are already built in. Millions of developers already know how to use its ecosystem. Hugging Face says that robotics is the category that is growing the fastest, and NVIDIA’s GR00T models are the most downloaded ones in that category. Already, companies like Boston Dynamics and NEURA Robotics are building on top of NVIDIA’s base.
Still, it’s not clear if one platform can really work for the huge range of tasks that robots will need to do in the future. There is a real and large difference between “picks up a cup in a controlled demo” and “handles a warehouse floor at 2 a.m. when something unexpected happens.” As I watch NVIDIA carefully put together this ecosystem of models, chips, simulation tools, academic partnerships, and the developer community, I get the sense that the infrastructure side of the robot boom may already be set, even though the robots themselves aren’t quite ready yet.
