NVIDIA Isaac GR00T, Explained: What NVIDIA Actually Sells in Robotics

01What Isaac GR00T is

Isaac GR00T is NVIDIA's foundation model for humanoid robots — a piece of software, not a physical product, meant to give a robot general-purpose perception, reasoning, and motor skills instead of hand-coded behavior for each task. NVIDIA first previewed the effort as "Project GR00T" at GTC in March 2024, then unveiled Isaac GR00T N1 a year later at GTC San Jose in March 2025, describing it in its own words as "the world's first open, fully customizable foundation model for generalized humanoid reasoning and skills."

The model uses a dual-system architecture: a fast, reflexive "System 1" that handles moment-to-moment motor control, paired with a slower, deliberate "System 2" vision-language model that plans and reasons about the task at hand — trained on a mix of egocentric human video, real robot trajectories, and synthetic data generated in simulation. CEO Jensen Huang framed the launch in sweeping terms: "The age of generalist robotics is here." NVIDIA's broader pitch to industrial customers, repeated across 2025 and 2026 announcements, goes further still: "every industrial company will become a robotics company."

02Development timeline (2024–2026)

GR00T's public history is a series of GTC-keynote announcements, each expanding either the model's capability or the compute it runs on, moving roughly twice a year since the March 2025 unveiling.

03The full stack

What NVIDIA actually sells is less a single product than a pipeline. Isaac Sim provides the physics simulation; Cosmos and Cosmos Transfer generate and photorealistically augment synthetic training video from that simulation; GR00T-Dreams, unveiled at Computex 2025, uses the same Cosmos world models to turn a single image and a language instruction into a synthetic robot-motion "dream" — NVIDIA says this cut the time to generate GR00T N1.5's training data to roughly 36 hours versus about three months of manual collection, though that specific figure is NVIDIA's own claim and hasn't been independently audited. Isaac Lab handles the reinforcement- and imitation-learning training itself; Isaac Teleop, which reached general availability at GTC 2026, lets developers capture real demonstrations in both simulation and the real world through one unified tool. GR00T is the foundation model that sits on top of all of it; Jetson AGX Thor is what a finished robot actually runs the model on.

04Ecosystem: who's actually building on it

Loose ties vs. confirmed ones

NVIDIA doesn't build humanoid robots itself — its business model is to sell the layers other companies build robots on top of, and a March 2026 press release naming dozens of partners at once is the clearest single snapshot of how that plays out in practice. A few of those relationships are concrete and independently verifiable, shown below. Others are looser: NVIDIA's own release groups eight humanoid makers (Agility Robotics, Boston Dynamics, Figure, 1X, AGIBOT, Hexagon Robotics, NEURA Robotics, and Humanoid) together as "building humanoids using Cosmos, Isaac Sim, Isaac Lab, and GR00T N models," without per-company specifics — language worth reading as directional rather than a confirmed deployment list for each name on it.

The Unitree reference platform

The most concrete recent example is the reference platform NVIDIA built with Unitree, unveiled at GTC Taipei in June 2026: a humanoid combining Unitree's "H2 Plus" chassis (roughly 1.8m, 31 degrees of freedom) with tactile five-finger hands from Singapore's Sharpa and a Jetson AGX Thor "brain" running the full Isaac software stack, for 75 degrees of freedom in total. NVIDIA named Ai2, ETH Zurich, Stanford, and UC San Diego as research-institution users. Coverage of the announcement noted NVIDIA choosing a Chinese robot maker for this specific reference design — though NVIDIA has said it is pursuing similar reference platforms with US and European humanoid makers too, so this isn't an exclusive arrangement.

Concrete, independently verifiable partnerships

Hugging Face and LeRobot

In July 2026, NVIDIA and Hugging Face expanded their partnership, integrating GR00T 1.7 and Isaac Teleop into the open-source LeRobot library and describing GR00T 1.7 as "the first open and commercially viable robot foundation model." The two companies claimed their combined ecosystems now connect NVIDIA's self-reported 3 million robotics developers with Hugging Face's 16 million AI builders — the kind of scale claim that's genuinely hard to verify from outside NVIDIA. See the linked article below for how that developer-count claim has moved over time.

05Scale next to the core business

Whatever the developer-adoption numbers mean, robotics doesn't even get its own disclosed revenue figure. NVIDIA's earnings commentary groups Automotive and Robotics under one heading, but the only dollar figure it discloses there is Automotive revenue: $2.3 billion for full fiscal year 2026, against $215.9 billion in total company revenue and $193.7 billion from Data Center alone — that $2.3 billion is an automotive number, not a robotics one (see the DRIVE article linked below for that business in detail). Robotics revenue on its own isn't broken out at all — automotive alone already comes to only about 1% of NVIDIA's business, next to Data Center's roughly 90%, so whatever robotics contributes specifically is smaller still. Starting with its Q1 FY2027 results, NVIDIA stopped reporting Automotive as a standalone line altogether, folding it — and robotics — into a broader "Edge Computing" category alongside PCs and game consoles, itself a sign of how small both lines remain relative to the rest of the company.

None of that means the bet is unserious — NVIDIA's market capitalization sat around $5.1 trillion as of mid-July 2026, the world's largest, built almost entirely on the data-center AI business the chart below tracks. Robotics is a long-run platform play layered on top of that, not something NVIDIA needs to work today.

06Challenges ahead

GR00T's code is released under an Apache 2.0 license, but the pretrained model weights on Hugging Face carry NVIDIA's own "Open Model License" — a standard type of license for pretrained weights that adds usage-policy terms for downstream commercial deployments, the same general pattern many other model providers use.

The competition is real, and not obviously behind. Google DeepMind's Gemini Robotics reportedly adapts to new tasks from 50–100 demonstrations using a single large model rather than GR00T's two-system split, though as of mid-2026 it remains limited to "selected trusted testers" rather than a public release — meaning it's arguably less open than GR00T even as it competes on capability. Physical Intelligence's π0 is a third approach again, and notably Physical Intelligence is simultaneously an NVIDIA compute customer (a named Jetson Thor evaluator) and a foundation-model competitor to GR00T — the industry's biggest names aren't cleanly sorted into "NVIDIA's ecosystem" versus "NVIDIA's rivals."

07FAQ

08Sources