NVIDIA Isaac GR00T N1

What it is

Isaac GR00T N1 is a foundation model for humanoid robots, announced by NVIDIA on 18 March 2025 at its GTC conference. NVIDIA described it as the first open, fully customizable foundation model for general-purpose humanoid reasoning and skills ^[1].

"Open" here does not mean a finished product. The intended way to use the model is to adapt it: developers and researchers can post-train GR00T N1 with their own real or synthetic data to fit a particular humanoid robot or a particular task ^[1]. A foundation model in this sense is a starting point — a broadly trained base that a robotics team can specialize rather than build from scratch.

How it works

GR00T N1 uses a dual-system architecture that NVIDIA says is inspired by the way human cognition is thought to work ^[1]. The two parts divide the job much as a slow, deliberate layer and a fast motor layer would.

System 2 is a vision-language model. It looks at the scene, takes in the instruction it has been given, reasons about what is in front of it and plans what should happen next ^[1]. System 1 takes those plans and converts them into the continuous, precise movements the robot needs in order to carry them out. It was trained on human demonstration data together with synthetic data generated by Omniverse, NVIDIA's simulation platform ^[1].

Splitting the work this way means the slower, general-purpose half can stay largely the same across tasks, while the fast half handles low-level control, which has to match a specific machine. The model is not tied to one task: NVIDIA lists grasping and moving objects with one or both arms, transferring items from one arm to the other, and multistep tasks that require long context among the things it can handle ^[1].

Where the training data came from

Robot policies are usually trained on human demonstrations, which are slow and expensive to collect and cover only a limited range of situations. NVIDIA's Isaac GR00T Blueprint, a pipeline for generating such data, produced 780,000 synthetic trajectories in 11 hours. NVIDIA says that volume is equivalent to 6,500 hours — nine continuous months — of human demonstration data ^[1].

Synthetic data was not used alone. NVIDIA reports that combining synthetic data with real data improved GR00T N1's performance by 40% compared with using only real data ^[1].

Early use and availability

In one demonstration, a humanoid robot from 1X autonomously performed domestic tidying tasks using a policy that had been post-trained on GR00T N1 ^[1]. NVIDIA named Agility Robotics, Boston Dynamics, Mentee Robotics and NEURA Robotics among the humanoid developers that received early access to the model ^[1].

NVIDIA also released GR00T N1's training data and task evaluation scenarios on Hugging Face and GitHub, so that outside teams can examine and reuse them ^[1].

Where it fits

GR00T N1 sits at the point where two lines of work meet: large vision-language models that can reason about images and instructions, and the low-level control software that makes a physical robot move. The dual-system design is an explicit attempt to combine them in one model that others can adapt.

For humanoid robotics, the notable part is not only the model itself but the data question around it. Generating synthetic trajectories in simulation, then mixing them with real demonstrations, is presented by NVIDIA as a way to get past the cost of collecting human demonstration data by hand, and the reported 40% gain is the evidence offered for it ^[1]. Because the model is released openly for post-training, its role is likely to be as a common base that different robot makers specialize in different directions, in the same way that the named early-access partners work on their own hardware.