NVIDIA Ships GR00T-H-N1.7, a Commercial Foundation Model for Surgical Robotics

NVIDIA has released GR00T-H-N1.7, the first commercially licensed foundation model built specifically for surgical and healthcare robotics, clearing a roadblock that has quietly throttled the field: until now, every team training a surgical robot had to assemble its own data, define its own action representations, and reconcile kinematics across incompatible hardware before it could even begin. The release, which landed in June 2026, upgrades the research-only GR00T-H that shipped at GTC in March with a new backbone, a broader training corpus, and — crucially — the NVIDIA Open Model License that makes production use legal for the first time.

The model is a roughly 3-billion-parameter vision-language-action (VLA) system. It reads camera frames and a natural-language description of a clinical task, then a diffusion-transformer action head generates the continuous motor commands needed to carry it out — suturing, tissue manipulation, endoscopy. It is built on the GR00T N1.7 humanoid backbone, which NVIDIA's VP of generative AI software Kari Briski described at GTC as part of an effort to "extend intelligence beyond language, enabling developers worldwide to build intelligent agents and power breakthroughs across digital and physical industries." GR00T-H-N1.7 is available now on Hugging Face and GitHub, and runs on NVIDIA's Isaac for Healthcare stack alongside the Cosmos-H surgical world model, Jetson AGX Thor compute, and the Holoscan platform.

The model sits on top of Open-H-Embodiment, billed as the largest open dataset for healthcare robotics assembled to date: 770 hours of CC-BY-4.0-licensed data spanning roughly 20 robot platforms and more than 50 institutions, including Johns Hopkins, Stanford, UC Berkeley, Vanderbilt, and Northwell Health, plus commercial robot makers CMR Surgical, Moon Surgical, Rob Surgical, and Virtual Incision. The surgical post-training subset used to produce GR00T-H-N1.7 is around 601 hours of real-world tasks across seven robot platforms. The single largest contributor is CMR Surgical, the Cambridge, UK maker of the Versius system, which supplied close to 500 hours of anonymized procedure data.

That data-for-tooling exchange is the commercial heart of the announcement. CMR is now using Cosmos-H to generate synthetic surgical video and evaluate new robot policies in simulation before any physical testing — NVIDIA reports that 600 rollouts that would take two days on a real benchtop run in about 40 minutes in Cosmos-H. "Surgical robotics generates a rich understanding of how procedures are performed," said Chris Fryer, CTO of CMR Surgical. "Because Versius is the most software-driven robot on the market, we were well-placed to share our data with the wider ecosystem... Combining clinical data with advances in AI and simulation creates a powerful opportunity to accelerate innovation responsibly."

Why a commercial license is the unlock

Foundation models reshaped software because a shared, pre-trained baseline shifted the work from building from scratch to fine-tuning. Surgical robotics never had that baseline — and the research-only license on the original GR00T-H meant that even where a model existed, it could not legally ship in a product. The Open Model License changes the calculus. Before, the company with the most high-quality teleoperated surgical video held a moat a rival could not buy overnight. Once a sufficient open foundation exists, the advantage migrates from raw data accumulation to fine-tuning effectiveness and, eventually, to who clears regulatory review fastest.

That is also why the partner roster matters more than the parameter count. CMR is simultaneously the most credentialed data contributor and an immediate beneficiary of the simulation tooling; Johnson & Johnson MedTech is using a Cosmos-based model for its MONARCH urology platform; and consortium partners are testing agentic orchestration of full OR workflows through the Rheo blueprint. NVIDIA's David Niewolny framed the bet plainly: medical-technology leaders training open models on the physical-AI platform are "accelerating a new generation of intelligent robotic systems that can assist surgeons, scale surgical expertise and ultimately expand access to high-quality care."

What to watch — and the hard limits

The enthusiasm should be read against the model card, which states that GR00T-H-N1.7 is not intended for clinical deployment, patient care, or medical decision-making. That is not boilerplate. On the SutureBot benchmark, GR00T-H is the first foundation model to complete a full suturing sequence end-to-end on a real surgical robot — but at a 25% success rate across 20 trials. That is a meaningful research milestone and nowhere near a clinical standard. Current systems sit at roughly level 2 autonomy (specific subtasks) in pre-clinical settings; the gap to routine use runs through prospective human trials, FDA clearance under the agency's 2025 draft guidance for AI-enabled devices, liability frameworks, and surgeon training standards — years of work, not months.

The pacing constraint is data. At 601 hours of surgical post-training, the corpus is large by the field's standards yet tiny next to the 20,000-plus hours of egocentric human video behind the base model. NVIDIA frames the current release as imitation learning from expert demonstrations, with a reasoning-capable next version of Open-H — annotated with intents, outcomes, and failure modes — as the real "ChatGPT moment" it is reaching for. What shipped in June is more modest and more useful: a commercially usable, technically specific starting point that did not exist six months ago. The thing to watch now is not autonomy in the OR, but how quickly partners turn that shared base into validated, fine-tuned policies — and how regulators respond when they do.