NVIDIA Isaac GR00T N1.7 overview and key facts (October 2026)

GR00T N1.7 is NVIDIA's open vision-language-action model for robot manipulation. Robotics teams at companies and research labs use it as a pretrained starting point and fine-tune it on their own robot and task data.

The model takes language and images as input and produces actions for manipulation tasks. It combines a vision-language backbone, in N1.7 based on Cosmos-Reason2-2B and Qwen3-VL, with a diffusion transformer head that generates continuous actions. The GitHub repository provides reference code for inference, fine-tuning and evaluation plus a LeRobot integration.

The GitHub README calls N1.7 a general availability release with support and stability guarantees, including production deployment with commercial support, while the Hugging Face model card labels it N1.7 EA (early access). The 3-billion-parameter checkpoint had about 135,000 downloads in the last 30 days on Hugging Face as of September 28, 2026. Earlier versions N1.5 and N1.6 remain available as branches.

• Turns instructions into robot actionsGiven a spoken or written instruction and camera images, the model predicts the arm and hand movements for tasks such as pick and place. Robotics teams can prototype new tasks by describing them instead of scripting every motion.
• Adapts to your robotFine-tuning on your own demonstrations or recorded data adapts the pretrained model to a specific robot and task. Teams need far fewer examples than training a policy from scratch, which shortens pilot projects.
• Covers several robot typesThe cross-embodiment training lets one model serve bimanual arms, semi-humanoids and humanoids. Companies testing different hardware can keep one software stack while they compare robots.

• The GitHub README describes a general availability release with a stated option for commercial support.
• Code under Apache 2.0; model card and README name the NVIDIA Open Model License for the weights, while the LICENSE file on Hugging Face is limited to research and evaluation.
• Works with the LeRobot data format, so datasets from that ecosystem can be reused.

GR00T is an open model from NVIDIA on Hugging Face and runs on your own hardware in any country.

GR00T is a model you download and run yourself, so your team needs NVIDIA GPUs for fine-tuning and inference plus the robot hardware. The license is contradictory: model card and README name the NVIDIA Open Model License, which permits commercial use, but the LICENSE file in the Hugging Face repository contains an NVIDIA license for research and evaluation only. Get written clarification from NVIDIA before production use. Robot actions come from a learned model, so plan safety testing and human supervision before the robot works near people.

No reviews yet. Be the first to share how your team uses it.

Every review shows its verification level. We read each one before it goes live. How we check reviews