nvidia/isaac-gr00t is the public repository NVIDIA publishes for general-purpose robot action models. It is written in Python and distributed under the Apache-2.0 license. The repository description presents the core of the project as GR00T N1.7, a foundation model for general robots including humanoids. The latest release is n1.7-release, and the repository covers the training and inference framework for general-purpose action models.
GR00T N1.7 is an embodied model that generates actions from vision, language, and robot state. The README introduces N1.7 as an open vision-language-action (VLA) model. Vision-language-action means a model that reads video and language and turns them into robot behavior. It takes multimodal inputs such as language and images, performs manipulation tasks in varied environments, and can be post-trained for a specific robot, task, or environment. Post-training is the step of re-aligning an already-trained model to a target condition. The workflow the README lays out runs from data preparation to inference, fine-tuning, evaluation, and deployment. The repository was created in March 2025 and has 8233 stars.
That said, the change list for n1.7-release is not organized. The README explains the training and inference flow but presents no benchmarks that back N1.7's performance with numbers. The Momentum score is also low at 3.6 out of 10. Judging the general-purpose performance now is difficult until post-training results for a specific robot and task are available.