During its GTC 2025 event, NVIDIA unveiled a technology portfolio intended to accelerate the development of humanoid robots. Among these innovations is Isaac Gr00T N1, the first open source and customizable foundation model, aimed at providing robots with generalized reasoning capacity and versatile skills, thus facilitating their adaptation to various tasks in various environments.

A new generation of humanoid robots

GR00T N1 marks a significant advance in generalist robotics by allowing developers to provide humanoids with advanced reasoning capacities. This model is based on a cognitive architecture inspired by human functioning combining two systems:

  • Vision-Language model: This methodical thinking system is based on Nvidia-Eagle with Smollm-1.7b. He interprets the environment through vision and language instructions, which allows robots to reason on their environment and their instructions, and to plan good deeds;

  • Diffusion transformer: This action model generates continuous actions to control the robot movements, translating the action plan developed by the methodical thinking system into precise and continuous robot movements.

GR00T N1 has been trained on a large set of data, including human demonstrations and synthetic data generated by the Omniverse Nvidia platform. This approach allows him to accomplish precise tasks such as gripping, manipulation of objects but also sequences of multi-stage action, with fluidity and precision.

NVIDIA has collaborated with leading companies such as Agility Robotics, Boston Dynamics, Mentoree Robotics, Neura Robotics and 1x Technologies that have benefited from early access to GR00T N1. During the opening Keynote, Jensen Huang, CEO of Nvidia, and Bernt Børnich, CEO of 1x technologies, demonstrated the capabilities of the model with the NEO Gamma robot, which did independently domestic storage tasks.

An improved simulation and learning ecosystem

At the same time, Nvidia has announced several complementary technologies:

  • Isaac gr00t blueprint : A framework for generating synthetic movement data for the training of humanoid robots;

  • Newton : An open source physical engine developed in collaboration with Google Deepmind and Disney Research which will be available in 2025. Built on Nvidia Warp, Newton is designed to optimize the learning of robots and will integrate with frameworks like Mujoco;

  • Mujoco-warp : Fruit of a collaboration between Google Deepmind and Nvidia, this project aims to accelerate robotic simulations of more than 70 times.

These advances strengthen the potential of humanoid robots, especially in industrial and entertainment applications. Disney Research plans to use Newton to develop more expressive robotic characters, like BDX droids inspired by Star Wars.

Jensen Huang said:

“The era of generalist robotics has arrived. With Nvidia Isaac Gr00T N1 and the new Frameworks of data generation and robotic learning, robotics developers around the world will open the next border in the AI ​​era.”

GR00T N1 training data and task assessment scenarios are available for download on Hugging Face And Github. The Nvidia Isaac Gr00t model for the generation of synthetic manipulation movement is also available in the form of an interactive demo on Build.nvidia.com or for download on Github.