NVIDIA announces Project GR00T

Background

Most robots in use today are built to do one job well, in a controlled setting, under specialist programming. Humanoid robots, with bodies shaped roughly like a person's, are meant to work in spaces designed for people, which makes broadly applicable skills more important.

A foundation model is a large model trained on broad data that can then be adapted to many tasks. Large language models are the best-known example. Applying the same idea to robots means training one model on a wide range of motion and sensor data, instead of writing separate control software for each task. Simulation is a common tool here: a robot can practise in a virtual world before it is tested in the physical one.

What happened

On 18 March 2024, at its GTC conference, NVIDIA announced Project GR00T, a general-purpose foundation model for humanoid robots.^[1] NVIDIA said robots powered by GR00T would understand natural language and learn skills such as coordination and dexterity by observing human actions.^[1] NVIDIA founder and CEO Jensen Huang described building foundation models for general humanoid robots as one of the most exciting problems to solve in AI.^[1]

The announcement came with hardware and software updates, which NVIDIA presented as the infrastructure for building such robots.

Key details

On the hardware side, NVIDIA announced Jetson Thor, a computer for humanoid robots based on the NVIDIA Thor system-on-a-chip. Its Blackwell-architecture GPU delivers 800 teraflops of 8-bit floating point AI performance.^[1] An 8-bit floating point format uses fewer bits per number, which speeds up AI computation at the cost of precision.

On the software side, the Isaac platform update added Isaac Lab, a GPU-accelerated reinforcement-learning application built on Isaac Sim that runs thousands of simulations in parallel. Reinforcement learning trains a program through trial, error and reward signals, and running many simulations at once lets a robot accumulate far more practice than a single physical machine could. The update also added OSMO, a compute orchestration service for generating robot data, training and testing.^[1]

NVIDIA named the humanoid makers it was working with: 1X Technologies, Agility Robotics, Apptronik, Boston Dynamics, Figure AI, Fourier Intelligence, Sanctuary AI, Unitree Robotics and XPENG Robotics.^[1]

A year later, at GTC on 18 March 2025, NVIDIA released Isaac GR00T N1. The company presented it as a fully customizable humanoid robot foundation model with two parts: a fast "System 1" action model, and a slower "System 2" planner based on a vision-language model, which handles images and language together.^[2] Early-access users included Agility Robotics, Boston Dynamics, Mentee Robotics and NEURA Robotics.^[2]

Impact

The 2024 announcement was largely a statement of direction, platform and tools. The 2025 release gave early users a concrete model to work with. Both were NVIDIA's own descriptions, and the performance figures cited here have not been independently verified.

The significance is that a chip and simulation-platform company moved into robot foundation models, carrying over the training approach that drove language models: broad pretraining followed by adaptation to specific tasks, with simulation supplying much of the practice data.