Repolode
Submit
NVIDIA

NVIDIA/Isaac-GR00T

NVIDIA's open vision-language-action foundation model (GR00T N1.7) for generalist humanoid and robot manipulation skills, with fine-tuning and deployment tooling.

View on GitHubWebsite
Stars
8,242
Forks
1,503
Open issues
346
Last commit
21 hours ago

What is it?

Isaac GR00T is NVIDIA's open vision-language-action (VLA) model for robots. It takes language instructions and camera images as input and outputs continuous robot actions, using a vision-language backbone combined with a diffusion transformer action head. The current N1.7 release uses a Cosmos-Reason2-2B (Qwen3-VL) backbone and is pretrained on robot demonstrations along with 20K hours of human video.

The repository covers the full workflow: converting demonstrations to the GR00T LeRobot-style dataset format, running zero-shot or finetuned inference, fine-tuning on a custom embodiment via a modality config, and evaluating with open-loop comparisons or closed-loop server-client rollouts. It includes finetuned checkpoints for benchmarks such as LIBERO, DROID, and SimplerEnv, plus deployment paths for ONNX and TensorRT on desktop GPUs and Jetson devices.

Model weights are released under the NVIDIA Open Model License, and the code is Apache 2.0. Inference requires a GPU with 16 GB+ VRAM, and the base model loads a gated Hugging Face backbone that needs access approval.

Key features

  • 01Vision-language-action model that maps language and images to robot actions
  • 02Fine-tuning on custom robot data through a modality config and launch script
  • 03Open-loop and closed-loop evaluation with a policy server and client
  • 04Pretrained checkpoints for LIBERO, DROID, and SimplerEnv benchmarks
  • 05ONNX and TensorRT export for deployment on desktop GPUs and Jetson devices
  • 06Dataset conversion tools for the GR00T LeRobot v2 format

Use cases

  • Fine-tuning a manipulation policy for a custom arm or humanoid using your own demonstrations
  • Running zero-shot inference with the base model on DROID-style robot data
  • Benchmarking a VLA policy on LIBERO or SimplerEnv simulation tasks
  • Deploying a robot policy to a Jetson or GPU server with TensorRT acceleration

Quick start

  1. 1

    Install git-lfs and clone the repository with submodules

    git clone --recurse-submodules https://github.com/NVIDIA/Isaac-GR00T
  2. 2

    Install uv (package manager)

  3. 3

    Create the environment and install GR00T dependencies

    uv sync --python 3.12
  4. 4

    Verify the installation

    uv run python -c "import gr00t; print('GR00T installed successfully')"

Things to know

  • Gated Hugging Face access is required for the Cosmos-Reason2-2B backbone before inference or fine-tuning.
  • Fine-tuning needs GPUs with 40 GB+ VRAM recommended, and inference needs 16 GB+ VRAM.
  • The default install needs FFmpeg 4-7 because torchcodec 0.8.0 does not support FFmpeg 8.

In collections

Robotics & embodied AI projects

Topics

#robotics#vision-language-action#foundation-model#imitation-learning#humanoid-robot#fine-tuning#tensorrt#embodied-ai

FAQ

What is Isaac-GR00T?
NVIDIA's open vision-language-action foundation model (GR00T N1.7) for generalist humanoid and robot manipulation skills, with fine-tuning and deployment tooling. Isaac GR00T is NVIDIA's open vision-language-action (VLA) model for robots. It takes language instructions and camera images as input and outputs continuous robot actions, using a vision-language backbone combined with a diffusion transformer action head. The current N1.7 release uses a Cosmos-Reason2-2B (Qwen3-VL) backbone and is pretrained on robot demonstrations along with 20K hours of human video.
Is Isaac-GR00T free to use? What license does it have?
Isaac-GR00T is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is Isaac-GR00T still maintained?
Actively maintained. The last commit to Isaac-GR00T was on Oct 21, 2026.
What can I use Isaac-GR00T for?
Fine-tuning a manipulation policy for a custom arm or humanoid using your own demonstrations Running zero-shot inference with the base model on DROID-style robot data Benchmarking a VLA policy on LIBERO or SimplerEnv simulation tasks Deploying a robot policy to a Jetson or GPU server with TensorRT acceleration
How do I install Isaac-GR00T?
Install git-lfs and clone the repository with submodules: git clone --recurse-submodules https://github.com/NVIDIA/Isaac-GR00T Install uv (package manager) Create the environment and install GR00T dependencies: uv sync --python 3.12 Verify the installation: uv run python -c "import gr00t; print('GR00T installed successfully')"
Where can I get Isaac-GR00T?
The source code and installation instructions are on GitHub: https://github.com/NVIDIA/Isaac-GR00T

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects