Project GR00T

What it is

Project GR00T is the name NVIDIA gave in March 2024 to an effort it described as a general-purpose foundation model for humanoid robots^[1]. A foundation model, in this setting, is a large model trained broadly enough to serve as a starting point for many downstream tasks, rather than a program written for one robot and one job. For humanoid machines — robots built to move and work in spaces designed for people — the appeal is that perception, language understanding and movement control could share one base model instead of being assembled separately for every new body.

Where it came from

Coverage of the announcement placed the project inside a wave of new humanoid robot companies. TechCrunch described NVIDIA as building an AI platform for the recent spate of entries into the category of humanoid robots^[1]. The same report listed the humanoid makers then active in the field: 1X Technologies, Agility Robotics, Apptronik, Boston Dynamics, Figure AI, Fourier Intelligence, Sanctuary AI, Unitree Robotics and XPENG Robotics^[1]. The list shows how fragmented the sector was: many companies, each with its own hardware, software stack and training pipeline. A shared platform would let those teams build on common tools instead of repeating the same groundwork.

GR00T N1

On 18 March 2025 NVIDIA announced Isaac GR00T N1, describing it as the world's first open, fully customizable foundation model for generalized humanoid reasoning and skills^[2]. Two words in that description carry weight. "Open" means the model is meant to be examined and adapted by outside teams rather than kept behind a service interface. "Fully customizable" means it is a starting point to be fine-tuned for a particular robot body and a particular task. NVIDIA said GR00T N1 is the first of a family of fully customizable models, so more members of that family were planned^[2].

NVIDIA also said the training data and task evaluation scenarios for GR00T N1 could be downloaded from Hugging Face and GitHub^[2]. Releasing data alongside a model is a practical choice: researchers can reproduce what is claimed, and robot makers can test a model against shared scenarios before committing to a hardware programme. Evaluation scenarios do a second job as well — they give the field a common yardstick. Without one, every lab reports results on its own machine and the numbers cannot be compared.

Where it sits

Humanoid robots are hard for reasons that are part mechanical and part computational. A machine that walks, grasps and avoids people in buildings designed for humans has to turn noisy streams from cameras and joint sensors into action, quickly and repeatedly; change the body or the task and much of that mapping has to be rebuilt. The foundation-model approach separates the shared part — understanding a scene, following an instruction — from the part that differs, and leaves the difference to fine-tuning.

That is why a common model, released together with data and evaluation scenarios, matters more than any single demonstration: it changes the unit of work from one robot at a time to a platform that many teams can use.