NVIDIA Isaac GR00T and Jetson Thor: the shared software and computer behind many humanoids

Isaac GR00T is a family of open AI models from NVIDIA that turn camera images and spoken or written instructions into robot movements. Jetson Thor is NVIDIA's computer module built to run models like this inside the robot. Together they form a shared starting point that many humanoid makers build on, which is why similar capabilities, and similar limits, appear across different brands.

Two products, one strategy

NVIDIA does not build humanoid robots. It sells the parts that sit underneath them: chips, simulation software and, since 2025, a ready-made robot "brain" that manufacturers can download and adapt. The strategy resembles what it did for self-driving cars: provide the computer and the training tools, and let others build the vehicle.

For anyone watching the humanoid market, this matters. When a new robot appears from a smaller manufacturer, part of what it can do may come from NVIDIA's stack rather than from the maker's own research. That is not a criticism; it is how much of the industry now works.

What Isaac GR00T is

NVIDIA announced Isaac GR00T N1 on 18 March 2025, describing it as an open and fully customisable foundation model for humanoid robots. A foundation model is a large AI model trained once on a broad mix of data, which developers then fine-tune for their own robot and tasks rather than training from scratch.

GR00T belongs to the category known as vision-language-action models, which we explain in more detail in our article on how vision-language-action models control robots. In plain terms, it takes in what the robot sees and what it has been told, and outputs a stream of motor commands for the arms and hands.

The "two systems" design

The GR00T N1 technical paper describes a dual-system architecture, loosely inspired by the idea of fast and slow thinking:

• System 2 is a vision-language model. It looks at the scene, reads the instruction and works out what needs to happen.
• System 1 is a diffusion transformer, a type of model that generates smooth sequences of actions. It converts that understanding into motor commands in real time.

The paper states that the two parts are trained together, end to end, on a mixture of real robot recordings, videos of humans and synthetically generated data. The researchers report testing it in simulation across several robot types and on a real Fourier GR-1 humanoid doing two-handed manipulation tasks guided by language.

Synthetic data as the shortcut

The biggest bottleneck in robot learning is data: recording a human operator guiding a robot through a task is slow and expensive. In its March 2025 announcement, NVIDIA said it generated 780,000 synthetic movement trajectories in 11 hours, roughly equal to nine months of continuous human demonstration, and that combining this synthetic data with real data improved GR00T N1's performance by 40% compared with real data alone. These are NVIDIA's own figures from its own tests.

How GR00T has changed since N1

NVIDIA has released updates at a steady pace. On 29 September 2025 it announced GR00T N1.6, saying it adds NVIDIA's Cosmos Reason model so that robots can break vague instructions into step-by-step plans, and that it lets humanoids move and handle objects at the same time.

NVIDIA's research team published technical details of GR00T N1.6 on 15 December 2025. According to that page, the action-generating part of the model doubled in depth compared with N1.5, and pretraining was extended with thousands of hours of data from bimanual arms, the AgiBot Genie-1 and the Unitree G1, among others. The team reports that N1.6 outperformed N1.5 in simulation and on those real robots.

At the time of writing, NVIDIA's Isaac-GR00T GitHub repository lists GR00T N1.7 as its current release. The repository gives it about 3 billion parameters, states that inference needs a single GPU with at least 16 GB of memory, and reports about 28 control updates per second on Jetson AGX Thor when optimised with NVIDIA's TensorRT software.

How open is "open"?

The code in the repository is published under the Apache 2.0 licence, while the model weights come under NVIDIA's own licence. The GR00T N1.7 model card on Hugging Face states that the model is ready for commercial and non-commercial use under the NVIDIA Open Model License. Licence terms have differed between releases, however, so a manufacturer has to check the exact terms of the version it uses. The model card also places responsibility on the developer to add guardrails and safety mechanisms before deployment.

What Jetson Thor is

A robot cannot rely on a distant data centre for split-second movement decisions; a dropped Wi-Fi connection would leave it frozen mid-step. So the model has to run on a computer inside the robot, within a limited power budget that competes with the motors for battery capacity.

NVIDIA made Jetson Thor generally available on 25 August 2025. It is based on the company's Blackwell GPU architecture. NVIDIA lists 2,070 FP4 teraflops of AI compute, 128 GB of memory and a 130-watt power envelope, and claims 7.5 times the AI compute and 3.5 times the energy efficiency of the previous Jetson Orin generation. In the same announcement, NVIDIA gave the developer kit a starting price of US$3,499.

The Jetson Thor product page lists two production modules:

• Jetson T5000: 2,070 TFLOPS (FP4, sparse), 128 GB of memory, a 14-core Arm CPU, 40 to 130 W.
• Jetson T4000: 1,200 TFLOPS (FP4, sparse), 64 GB of memory, a 12-core Arm CPU, 40 to 70 W.

The headline teraflops figures use a low-precision number format (FP4) with sparsity, which flatters peak numbers; real-world speed depends on the model and how it is optimised. The roughly 28 updates per second that NVIDIA reports for GR00T N1.7 on Jetson AGX Thor is a more practical indicator.

Who uses them

NVIDIA names many robot makers in its announcements, but the lists describe different levels of involvement, and they should be read carefully:

• GR00T early access: in March 2025 NVIDIA listed 1X Technologies, Agility Robotics, Boston Dynamics, Mentee Robotics and NEURA Robotics.
• Jetson Thor adoption: in August 2025 NVIDIA said Agility Robotics, Amazon Robotics, Boston Dynamics, Caterpillar, Figure, Hexagon, Medtronic and Meta were among those adopting Thor, with 1X, John Deere, OpenAI and Physical Intelligence evaluating it.
• Further Thor partners: in September 2025 NVIDIA added Galbot, Google DeepMind, Mentee Robotics, Skild AI and Unitree to the list of companies building on Thor.

Being named in an NVIDIA press release does not mean a company's production robot runs GR00T. Several makers, including Figure, develop their own AI models and may use NVIDIA hardware or simulation tools without using GR00T itself. Unless a manufacturer states which model and computer a specific robot uses, it is safer not to assume.

What GR00T and Thor do not solve

A shared model lowers the cost of getting a humanoid to pick things up and follow instructions, but it does not make the robot reliable in every setting. The published results come from controlled tasks such as folding, table clearing and placing objects, measured by NVIDIA's own researchers. Performance in a crowded, noisy, brightly lit event hall is a different test.

Balance and walking are also often handled by separate controllers, and the safety of a robot near people depends on the whole system: its mechanics, its stop functions, its software and the operator. Neither GR00T nor Thor is a safety certification. Many robots at events are also partly or fully remote-controlled, a distinction we explain in our guide to teleoperated and autonomous robots.

What this means for event organisers

• Expect faster improvement across brands. Because many makers can start from the same open model and computer, new manipulation skills may spread across several robots quickly, including from smaller manufacturers.
• Ask what actually runs the robot. If a supplier describes a robot as AI-powered, it is reasonable to ask which model and onboard computer it uses, and which parts of the act are autonomous and which are operated by a person. Our guide on briefing a robot supplier covers what to include.
• Treat lab results as lab results. Success rates published by NVIDIA come from specific tasks under test conditions; they do not predict performance at a busy stand.
• Safety still sits with the whole system. A capable AI model does not replace barriers, stop procedures and trained staff, as set out in our planning guide to safety at public events.

Frequently asked questions

Is NVIDIA Isaac GR00T a robot?

No. GR00T is a family of AI models, not a physical robot. NVIDIA publishes the models so that robot manufacturers can download them, fine-tune them with data from their own machines and use them to turn camera images and instructions into arm and hand movements. The robot itself, including its motors, sensors, safety functions and walking controller, comes from the manufacturer.

Is GR00T free to use commercially?

The code is published under the Apache 2.0 licence. The GR00T N1.7 model card on Hugging Face states that the model weights are released under the NVIDIA Open Model License and are ready for commercial and non-commercial use. Licence terms have differed between versions, so developers need to check the exact licence for the release they use. This is general information, not legal advice.

How much does Jetson Thor cost?

When NVIDIA announced general availability on 25 August 2025, it gave the Jetson AGX Thor developer kit a starting price of US$3,499. Production modules, the Jetson T5000 and T4000, are sold through distribution partners, and NVIDIA's product page does not list their prices. The cost of the computer is only a small part of the price of a complete humanoid robot.

Does every humanoid that uses Jetson Thor run GR00T?

No. Jetson Thor is a general-purpose robot computer and can run many different AI models. Several companies named by NVIDIA as Thor adopters, such as Figure, develop their own models. Equally, GR00T can run on other NVIDIA hardware. Unless a manufacturer states which model a particular robot uses, it is best not to assume a link between the two products.

How this article was made: written by the humanoid4rent editorial team with AI assistance; every factual statement was checked against the sources listed below on 2026-10-31. Spotted an error? Write to info@humanoid4rent.com.