Isaac GR00T N1.7 is NVIDIA’s open robot foundation model: a 3-billion-parameter vision-language-action (VLA) model that takes camera images, the robot’s joint state and a text instruction, and outputs continuous motor commands. The code is Apache 2.0, the weights are under the NVIDIA Open Model License, which allows commercial use, and it runs on a single GPU with 16 GB or more of memory. Fine-tuning needs about 40 GB. GR00T N2 was previewed in March 2026 but hasn’t been released as of October 23, 2026.
How GR00T N1.7 works
GR00T keeps the two-part design NVIDIA introduced with N1 in March 2025. A vision-language model (System 2) reads the camera images and the instruction and works out what to do. An action model (System 1), a diffusion transformer, turns that plan and the robot’s current state into smooth joint commands. In N1.7 the vision-language part is built on Cosmos-Reason2-2B, and the model was pre-trained on 20,854 hours of first-person human video, according to NVIDIA’s Hugging Face introduction (April 17, 2026).
The model card lists the inputs as one or more RGB frames, a state vector, a text instruction and an embodiment ID, which tells the model which robot it’s controlling. The GitHub repository gives an action horizon of 40 steps. NVIDIA validated N1.7 on the Unitree G1 humanoid, YAM bimanual arms and AgiBot Genie 1, and the repository includes examples for the LIBERO, DROID and SimplerEnv benchmarks and an SO-100 arm.
Version history
| Version | Announced | Notes |
|---|---|---|
| N1 | March 18, 2025 (GTC) | First open release, weights the same day |
| N1.5 | May 18, 2025 (Computex) | Weights on Hugging Face afterward |
| N1.6 | September 29, 2025 | Uses Cosmos Reason |
| N1.7 | March 16, 2026 (GTC) | Early access with commercial license; weights on Hugging Face |
| N2 | Previewed March 16, 2026 | Promised by the end of 2026; not released yet |
Sources: NVIDIA’s N1 release, N1.6 announcement and GTC 2026 release. The GitHub README now calls N1.7 generally available, while NVIDIA’s developer page still describes GR00T as early access; for a product, treat it as a model you validate yourself.
The license, in plain terms
- Code: Apache 2.0, per the repository.
- Weights: the NVIDIA Open Model License. It allows commercial use without royalties. If you redistribute the model, you must include the notice “Licensed by NVIDIA Corporation under the NVIDIA Open Model License.” Your rights end if you remove the model’s safety guardrails without an equivalent replacement, or if you sue NVIDIA over the model.
- Backbone access: the Cosmos-Reason2-2B backbone is gated on Hugging Face, so you need to log in and request access before the first run.
Some NVIDIA pages describe N1.7 as “Apache 2.0” overall; the weights license above is the one that governs the model file.
Hardware: what GPU you need
The repository recommends one GPU with 16 GB+ for inference, such as an RTX 4090, L40, H100, Jetson AGX Thor or Orin, or DGX Spark, and 40 GB+ (H100 or L40 preferred) for fine-tuning. Official latency with four denoising steps and one camera, from the model card:
| Device | Mode | Latency | Control rate |
|---|---|---|---|
| H100 | PyTorch | 85.8 ms | 11.7 Hz |
| H100 | TensorRT | 27.9 ms | 35.9 Hz |
| Jetson AGX Thor | TensorRT | 93.8 ms | 10.7 Hz |
| Jetson Orin | TensorRT (action model only) | 216.5 ms | 4.6 Hz |
For a robot that must run GR00T onboard, that makes Jetson AGX Thor the practical choice; our Jetson Thor vs Orin comparison covers price and specs. To fine-tune at a desk, DGX Spark has the memory, though it trains more slowly than an H100.
How to run GR00T N1.7
From the repository README:
- Install git-lfs, then clone with submodules:
git clone --recurse-submodules https://github.com/NVIDIA/Isaac-GR00T. - Install uv and FFmpeg 4 to 7, then run
uv sync --python 3.12. Separate install scripts cover DGX Spark, Jetson Thor and Orin, and there’s a Docker image. - Log in to Hugging Face and request access to Cosmos-Reason2-2B.
- Run the standalone inference script on the bundled DROID sample with
--model-path nvidia/GR00T-N1.7-3B; the roughly 6 GB checkpoint downloads on first use. A server-client mode lets a robot computer call the model over the network. - To fine-tune for your robot, record demonstrations in LeRobot v2 format with a
meta/modality.jsonfile, then run the fine-tuning launcher with theNEW_EMBODIMENTtag.
NVIDIA doesn’t give an official number of demonstrations to record. Its examples range from 50 episodes in the LeRobot walkthrough to 400 demonstrations in NVIDIA’s own end-to-end blog. Our LeRobot getting-started guide covers recording data, and the SO-101 kit guide covers a low-cost arm to practice on.
GR00T against other robot models
GR00T is the open option aimed at humanoids. OpenVLA is a smaller open alternative for arms, Physical Intelligence’s π models are open through openpi, and Google’s Gemini Robotics VLA is limited to partners. For where GR00T sits among NVIDIA’s other tools (Cosmos, Isaac Sim, Isaac Lab, Jetson Thor), see our NVIDIA physical AI guide.
NVIDIA repository, model card and news pages checked October 23, 2026.