In his keynote speech on the GTC 2025 convention, NVIDIA CEO Jensen Huang introduced an enormous leap within the growth of humanoid robots: the NVIDIA Isaac GR00T N1.
The N1 mannequin is the inaugural basis mannequin launched inside NVIDIA’s GR00T general-purpose robotic growth platform. It is usually the “world’s first open mannequin” for humanoid robotics.
NVIDIA GR00T N1 mannequin structure
It’s a totally customizable mannequin that makes use of a dual-system structure impressed by human cognitive processing. The primary system is a “fast-thinking motion mannequin” that allows exact and steady robotic actions, and the second is a “slow-thinking” planning mannequin that helps the robotic make deliberate selections about its setting and directions. Each techniques are built-in tightly, and adjustments made to the mannequin are utilized holistically.
In his announcement, Huang acknowledged, “By the top of this decade, the world goes to be no less than 50 million staff brief,” earlier than emphasizing the necessity for humanoid robots to fill the labor hole.
NVIDIA had earlier introduced Undertaking GR00T, a sophisticated growth platform for next-generation humanoid robotics constructed on the foundational robotic studying framework, Isaac Lab. Alongside pre-trained basis fashions for robotic “cognition and management,” GR00T affords simulation-ready 3D belongings, simulation frameworks, and artificial knowledge blueprints, all working on the purpose-built Jetson AGX Thor.
GR00T makes it simpler and cheaper to coach humanoid robots by leveraging human demonstration knowledge, artificial knowledge, and optimized workflows constructed on the NVIDIA Omniverse and Cosmos platforms.
NVIDIA Omniverse is a “digital twin working system” that gives builders with the constructing blocks for creating “bodily correct world-scale simulations” for functions equivalent to robotics, autonomous automobiles, and imaginative and prescient AI. Cosmos is a “world basis mannequin growth platform” for creating fashions able to understanding and producing lifelike representations of the true world. In less complicated phrases, Omniverse emphasizes bodily accuracy and Cosmos focuses on photorealism.
Actual-world sensor and/or demonstration knowledge is aggregated in Omniverse to generate rendered picture or video outputs, that are then used alongside textual content prompts to situation Cosmos. This permits Cosmos to create a just about infinite variety of managed, photoreal, and various artificial knowledge for bodily AI coaching.
NVIDIA Isaac GR00T Blueprint structure
The GR00T N1 mannequin was skilled on an “expansive humanoid dataset,” combining robotic knowledge obtained by way of teleoperation, artificial knowledge created utilizing the GR00T Blueprint, and human movies from the Web. NVIDIA says the combination of actual and artificial knowledge “resulted in a 40% efficiency enhance for GR00T N1 in comparison with utilizing solely actual knowledge.”
N1 will be fine-tuned for particular embodiments, duties, and environments utilizing a customized dataset. As a part of the keynote, Jensen showcased an autonomy demo that includes 1X Applied sciences’ humanoid robotic, NEO, loading a dishwasher with a post-trained coverage constructed on GR00T N1. 1X CEO Bernt Børnich mentioned, “Whereas we develop our personal fashions, NVIDIA’s GR00T N1 offers a major enhance to robotic reasoning and expertise.”
The GR00T N1 mannequin has additionally been built-in into the Fourier GR-1 humanoid robotic, which executes a collaborative multi-step sequence — greedy, transferring, and passing — within the NVIDIA GTC demo. The GR00T Blueprint, made up of particular workflows operated throughout the open-source Isaac Lab framework, has been tailored by humanoid builders equivalent to Agibot, Mentee Robotics, UCR, and X-Humanoid.
The GR00T platform is predicted to profit different main robotics firms like UNITREE, Determine AI, NEURA Robotics, Galbot, Agility Robotics, Sanctuary AI, Boston Dynamics, and Apptronik, amongst others. A few of these firms have acknowledged that they’re engaged on incorporating the GR00T N1 mannequin into their current humanoid robots.
Though N1’s core focus is humanoids, Staff Firebreathing Rubber Duckies fine-tuned the mannequin to run on two LeRobot SO-100 robotic arms for a bimanual pick-and-place activity. The undertaking gained first place within the Embodied AI Hackathon.
Huang declared that the “age of generalist robots is right here,” however the use circumstances for humanoid robots are nonetheless pretty restricted and simply replaceable by specialist robots. Generalist robots might change into widespread and drive a booming trade, however solely as soon as they surpass specialist robots in usefulness.
NVIDIA is, nevertheless, bullish on the robotics market, and GR00T N1 is barely the primary in a collection of customizable fashions that the corporate will “pretrain and launch to worldwide robotics builders.”
The GR00T N1 mannequin will be downloaded from Hugging Face, together with a subset of the coaching knowledge. The Isaac-GR00T GitHub repository offers pattern datasets and PyTorch scripts for fine-tuning. The GR00T Blueprint for artificial movement era is out there on GitHub and as an interactive demo on construct.nvidia.com.
