NVIDIA Isaac GROOT N1 Is an Open Source Foundation Model for Accelerated Humanoid Robot Development

In his keynote speech at the GTC 2025 conference, NVIDIA CEO Jensen Huang announced a massive leap in the development of humanoid robots: the NVIDIA Isaac GR00T N1.

The N1 model is the inaugural foundation model introduced within NVIDIA’s GR00T general-purpose robot development platform. It is also the “world’s first open model” for humanoid robotics.

It is a fully customizable model that uses a dual-system architecture inspired by human cognitive processing. The first system is a “fast-thinking action model” that enables precise and continuous robot actions, and the second is a “slow-thinking” planning model that helps the robot make deliberate decisions about its environment and instructions. Both systems are integrated tightly, and changes made to the model are applied holistically.

In his announcement, Huang stated, “By the end of this decade, the world is going to be at least 50 million workers short,” before emphasizing the need for humanoid robots to fill the labor gap.

NVIDIA had earlier announced Project GR00T, an advanced development platform for next-generation humanoid robotics built on the foundational robot learning framework, Isaac Lab. Alongside pre-trained foundation models for robotic “cognition and control," GR00T offers simulation-ready 3D assets, simulation frameworks, and synthetic data blueprints, all running on the purpose-built Jetson AGX Thor.

GR00T makes it easier and cheaper to train humanoid robots by leveraging human demonstration data, synthetic data, and optimized workflows built on the NVIDIA Omniverse and Cosmos platforms.

NVIDIA Omniverse is a “digital twin operating system” that provides developers with the building blocks for creating “physically accurate world-scale simulations” for applications such as robotics, autonomous vehicles, and vision AI. Cosmos is a “world foundation model development platform” for creating models capable of understanding and generating realistic representations of the real world. In simpler terms, Omniverse emphasizes physical accuracy and Cosmos focuses on photorealism.

Real-world sensor and/or demonstration data is aggregated in Omniverse to generate rendered image or video outputs, which are then used alongside text prompts to condition Cosmos. This enables Cosmos to create a virtually infinite number of controlled, photoreal, and diverse synthetic data for physical AI training.

The GR00T N1 model was trained on an “expansive humanoid dataset,” combining robot data obtained via teleoperation, synthetic data created using the GR00T Blueprint, and human videos from the Internet. NVIDIA says the integration of real and synthetic data “resulted in a 40% performance boost for GR00T N1 compared to using only real data.”

N1 can be fine-tuned for specific embodiments, tasks, and environments using a custom dataset. As part of the keynote, Jensen showcased an autonomy demo featuring 1X Technologies’ humanoid robot, NEO, loading a dishwasher with a post-trained policy built on GR00T N1. 1X CEO Bernt Børnich said, “While we develop our own models, NVIDIA’s GR00T N1 provides a significant boost to robot reasoning and skills.”

The GR00T N1 model has also been integrated into the Fourier GR-1 humanoid robot, which executes a collaborative multi-step sequence — grasping, transferring, and passing — in the NVIDIA GTC demo. The GR00T Blueprint, made up of specific workflows operated within the open-source Isaac Lab framework, has been adapted by humanoid developers such as Agibot, Mentee Robotics, UCR, and X-Humanoid.

The GR00T platform is expected to benefit other leading robotics companies like UNITREE, Figure AI, NEURA Robotics, Galbot, Agility Robotics, Sanctuary AI, Boston Dynamics, and Apptronik, among others. Some of these companies have stated that they are working on incorporating the GR00T N1 model into their existing humanoid robots.

Although N1’s core focus is humanoids, Team Firebreathing Rubber Duckies fine-tuned the model to run on two LeRobot SO-100 robot arms for a bimanual pick-and-place task. The project won first place in the Embodied AI Hackathon.

Huang declared that the “age of generalist robots is here,” but the use cases for humanoid robots are still fairly limited and easily replaceable by specialist robots. Generalist robots could become widespread and drive a booming industry, but only once they surpass specialist robots in usefulness.

NVIDIA is, however, bullish on the robotics market, and GR00T N1 is only the first in a series of customizable models that the company will “pretrain and release to worldwide robotics developers.”

The GR00T N1 model can be downloaded from Hugging Face, along with a subset of the training data. The Isaac-GR00T GitHub repository provides sample datasets and PyTorch scripts for fine-tuning. The GR00T Blueprint for synthetic motion generation is available on GitHub and as an interactive demo on build.nvidia.com.