Share
Jim Fan, who co-leads Nvidia’s Project GR00T, has confirmed that the teaser for his upcoming keynote shows real humanoid hardware operating autonomously at normal speed.
That clarification matters more than it sounds. Any robot video now gets asked the same questions: is it sped up, and is someone driving it from behind the curtain? Fan’s answer to both is no.
The keynote is scheduled for the Conference on Robot Learning (CoRL) 2026, set for November 9, 2026, in Austin, Texas. Fan is using the stage to show where Nvidia’s humanoid robotics effort has landed.
What Fan confirmed, and why it counts
Fan’s statement covers three things at once. The hardware is physical rather than simulated. The robot runs autonomously rather than under teleoperation. And the footage plays at its actual speed.
What Project GR00T is trying to build
Project GR00T is Nvidia’s push to create foundation models for humanoid robots. These are general-purpose AI systems that combine vision, language, and action in a single model.
Nvidia CEO Jensen Huang publicly unveiled the project at GTC 2024. The name stands for “Generalist Robot 00 Technology.”
The model releases have moved quickly since then. Nvidia introduced the open-weight GR00T N1 in March 2025, a 2B-parameter dual-system VLA, short for vision-language-action model.
Follow-up versions N1.5 and N1.6 add integration with Cosmos Reason. Those iterations were expected to roll out through 2025.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
We respect your privacy. Unsubscribe anytime.
The data problem at the center of robotics
The core obstacle GR00T targets is data scarcity. Large language models learned from enormous volumes of text. Robots have no equivalent internet-sized library of physical movement to train on.
Collecting real robot data is slow and expensive. It typically requires teleoperation, where a human controls the robot so the system can record what successful actions look like.
Nvidia’s answer is to multiply that scarce real data with simulation. The project uses platforms such as Isaac Lab and Cosmos, along with advanced teleoperation methods and large-scale synthetic data generation.
The headline number is striking. The methodology expands limited teleoperation data by over 1,000x, using simulations and generative world models to add variety in visuals and motion.
A pattern of CoRL showcases
This is not Nvidia’s first time using CoRL to show its robotics work. CoRL 2025 featured GR00T updates alongside a beta of the Newton physics engine.
The 2026 edition continues that thread, with a focus on advancing world models in robotics. World models are AI systems that learn how the physical world behaves, so a robot can anticipate outcomes before it acts.
What this means
For Nvidia, GR00T fits a broader strategy of supplying the tools other robot builders depend on. The company offers foundation models, simulation platforms like Isaac Lab and Cosmos, and physics engines like Newton.
Releasing GR00T N1 with open weights also means outside developers can test and build on the models.
The project’s stated aim is versatile robots that can handle multiple tasks across varied scenarios, not a single trick performed well.
The things to watch on November 9 are the range of tasks the robots handle, how the training pipeline is described, and how much of the result traces back to the simulation pipeline versus real-world data.
