When NVIDIA says “physical AI,” it means a stack of products for building robots and vehicles that act in the real world, organized as three computers: DGX systems to train the AI, Omniverse and Cosmos on RTX PRO servers to simulate and test it, and Jetson AGX Thor to run it on the robot. For what physical AI is already doing outside NVIDIA’s labs, see our physical AI examples. That split comes from NVIDIA’s own three-computer explainer. Omniverse, Isaac Sim, Cosmos and GR00T can be used free for internal research and development, so the basic cost is hardware such as Jetson Thor and DGX. Redistributing Isaac Sim, offering it as a service or installing it on a client’s hardware needs a paid NVIDIA AI Enterprise license.
NVIDIA’s glossary defines physical AI as AI that lets autonomous machines perceive, understand and perform complex actions in the physical world. For the general idea beyond NVIDIA’s products, start with our guide to what physical AI is.
The NVIDIA physical AI stack at a glance
| Part | What it does | Current version (Oct 2026) | Cost or license |
|---|---|---|---|
| Cosmos 3 | World model: generates, predicts and reasons about video of physical scenes | Cosmos 3 Super and Nano (Edge coming) | Open weights, OpenMDW 1.1 |
| Isaac GR00T | Robot foundation model that turns camera images and instructions into humanoid actions | GR00T N1.7 (N2 previewed) | Open weights, NVIDIA Open Model License |
| Isaac Sim | Robot simulator built on Omniverse | 6.2.0 | Free for R&D; redistribution needs a license |
| Isaac Lab | Training framework for robot policies in simulation | 3.0 early access | Open source |
| Newton | Open physics engine for robot learning | 1.x | Linux Foundation project |
| Omniverse | OpenUSD-based platform for digital twins and simulation apps | Kit SDK | Free for development and production |
| Jetson AGX Thor | Computer inside the robot | T5000 module; T4000 smaller module | Paid hardware |
| DGX Spark / DGX | Training and fine-tuning machines | DGX Spark 128GB | Paid hardware |
Train: DGX and DGX Spark
Training robot models takes large GPU clusters, which NVIDIA sells as DGX systems and rents through cloud partners. For a single developer, the desk-sized DGX Spark (128GB unified memory, 273GB/s) is enough to fine-tune mid-size models such as GR00T or run Cosmos inference locally.
Simulate: Omniverse, Isaac Sim, Isaac Lab, Newton and Cosmos
This is where most of the stack lives. What’s free depends on the product and how you use it.
- Omniverse is the base layer: an OpenUSD scene format plus the Kit SDK for building simulation and digital-twin apps. NVIDIA’s Omniverse page says that since May 2026 it is free for both development and production; an NVIDIA AI Enterprise subscription only adds paid support.
- Isaac Sim is NVIDIA’s robot simulator on top of Omniverse. Its license FAQ says it is free for internal R&D, and you can sell what you build with it; you need an NVIDIA AI Enterprise license to redistribute Isaac Sim itself, offer it as a service to third parties or ship it inside a turnkey system. The FAQ doesn’t address internal production use outside R&D either way, so check the license text or ask NVIDIA. The release notes list 6.2.0 as the latest version.
- Isaac Lab trains robot policies with reinforcement and imitation learning inside Isaac Sim. Version 3.0 is in early access, with general release targeted for late October 2026; our Isaac Lab getting-started guide covers which version to install.
- Newton is an open-source physics engine built by NVIDIA, Google DeepMind and Disney Research and contributed to the Linux Foundation in September 2025. Isaac Sim 6.x and Isaac Lab 3.0 use it alongside PhysX.
- Cosmos is NVIDIA’s world foundation model family. Cosmos 3, launched at GTC Taipei at the start of June 2026, folds the old Predict, Transfer and Reason models into one model that can generate synthetic video, reason about a scene and output actions. Robotics users named at launch include Doosan Robotics, LG Electronics and Samsung Electronics. Our Cosmos guide covers sizes and GPU requirements.
Run on the robot: GR00T and Jetson Thor
Isaac GR00T is the model that actually drives a humanoid’s arms and hands. The GR00T repository lists N1.7 as generally available, with 3B-parameter weights on Hugging Face for commercial and non-commercial use. At GTC in March 2026 NVIDIA previewed GR00T N2 for release by the end of 2026; it hasn’t shipped as of October 23. Our GR00T N1.7 guide covers the license, GPU needs and setup steps. GR00T is one of several robot foundation models; OpenVLA, Physical Intelligence’s π models and Google’s Gemini Robotics compete for the same job, and our OpenVLA beginner’s guide shows what an open alternative needs in hardware.
Jetson AGX Thor is the computer that runs those models inside the robot. The T5000 module has a Blackwell GPU rated at 2,070 FP4 TFLOPS, 128GB of memory and a 40–130W power range, per NVIDIA’s launch release. The developer kit launched at $3,499 in August 2025 and now lists at $5,499 after NVIDIA’s July 2026 price increase; our Jetson Thor vs Orin price comparison tracks the current numbers. A smaller T4000 module (64GB, 40–70W) is available now, and T3000 and T2000 modules are due in the first quarter of 2027, according to NVIDIA’s July 15 announcement. If your robot runs smaller vision models, the older Jetson Orin Nano and Orin NX still cost far less.
For vehicles the same pattern uses different parts: DRIVE AGX Thor in the car and the Alpamayo reasoning model, whose open weights NVIDIA released in January 2026 with commercial-use options promised for later versions.
Who builds on it
NVIDIA’s GTC 2026 release names ABB, FANUC, KUKA, Yaskawa and Universal Robots among industrial partners, and Agility, Boston Dynamics, Figure, 1X, AgiBot and NEURA among humanoid and general-purpose robot makers. “Partner” in these releases covers anything from using Isaac Sim for testing to putting Jetson Thor in a shipping product, so check each company’s own announcement for what it actually uses. For a map of who builds models and who builds bodies, see our list of humanoid robot companies and the Universal Robots g-Series, which ships with ROS 2 and open APIs aimed at this kind of AI integration.
Where to start if you want to try it
- Install Isaac Sim and Isaac Lab on a Linux PC with an RTX GPU; both are free for development.
- Pull a Cosmos 3 or GR00T N1.7 checkpoint from Hugging Face to see what the models produce before buying robot hardware.
- Buy a Jetson Thor developer kit only when you have a policy that needs to run on a real robot.
For self-driving, NVIDIA’s open reasoning models are covered in NVIDIA Alpamayo explained: versions, GPU needs and the OpenMDW license.
NVIDIA product, license and news pages checked October 23, 2026.