RoboBoom
AI Engineer· July 28, 2025· 17:47

What Is a Humanoid Foundation Model? An Introduction to GR00T N1 - Annika & Aastha

Annika and Aastha, NVIDIA team members who helped develop the GR00T N1 robotics foundation model, explain its data strategy, architecture, training, and use across robot embodiments at an AI Engineer event.

In this talk

  1. 0:31

    Why humanoid robots

    Annika points to a McKinsey Global Institute report describing a persistent gap between available jobs and people able to fill them. She highlights healthcare, construction, transportation, hospitality, and manufacturing as fields where work involves physical devices and cannot be handled by language models alone. A human-like form can help a general-purpose robot work in spaces built for people, unlike a specialist such as a barista robot that cannot switch to cooking rice.

  2. 2:38

    The physical AI life cycle and data pyramid

    Annika describes a three-stage process: generate data, train a model, and deploy it on a robot. She calls it a three-computer problem: simulation uses systems such as OVX Omniverse, training uses DGX, and deployment uses an edge device such as AGX. GR00T N1 is part of NVIDIA’s broader Project GR00T strategy; the two-billion-parameter model is open source and intended for customization across embodiments. Robot data is scarce and costly, so the team combines limited real teleoperation data with internet videos and synthetic data, including trajectories expanded with video generation models such as DreamGen.

  3. 7:49

    How GR00T N1 maps inputs to robot actions

    Aastha describes the model’s inputs as images, robot state, and a language instruction, with an action trajectory as output. GR00T N1 uses a dual-system design inspired by Daniel Kahneman’s *Thinking, Fast and Slow*: System 2 plans complex tasks, while System 1 executes quickly at nearly 120 hertz. State and action encoders, a diffusion transformer, and NVIDIA’s Eagle 2 vision-language model process inputs; an embodiment-specific action decoder turns the result into robot motion. That decoder lets the model’s learned knowledge be adapted to humanoid hands, robot arms, and industrial arms.

  4. 13:07

    Training and adapting a generalist model

    The speakers contrast imitation learning, which trains a robot to match expert demonstrations, with reinforcement learning, which improves behavior by maximizing reward through trial and error. Imitation learning is limited by the cost and availability of expert data, while reinforcement learning faces the challenge of transferring behavior from simulation to reality. GR00T N1 uses both approaches and is presented as adaptable to varied tasks and environments. Aastha concludes that co-training System 1 and System 2 helps optimize the overall system, while the generalist model can be fine-tuned for different embodiments and downstream tasks.

Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.

More talks