Nvidia's healthcare robotics stack is six months old this week, and it is worth being precise about what it is: not a product, not a robot, and not anything a patient will encounter. On March 16, 2026, at GTC in San Jose, the company released four things at once — Open-H, a 778-hour open dataset of surgical robotics data; Cosmos-H, a set of generative models that produce synthetic surgical video; GR00T-H, a vision-language-action model for clinical tasks; and Rheo, a blueprint for simulating an entire hospital. Half a year on, the release has held up better than most GTC announcements, mostly because the hard part it addressed was never the model.
The hard part was data. Surgical robotics has no ImageNet. Every team building a robot policy assembled its own corpus, defined its own action representations, and reconciled kinematics across hardware that does not agree on coordinate conventions, camera placement, or control frequency. Open-H-Embodiment was an attempt to end that duplication by fiat. Started by a steering committee of Prof. Axel Krieger at Johns Hopkins, Prof. Nassir Navab at the Technical University of Munich and Dr. Mahdi Azizian at Nvidia, it grew to 35 contributing organizations — Northwell Health, Stanford, UC Berkeley, Vanderbilt, Moon Surgical, Virtual Incision and a long tail of university hospitals — and shipped 778 hours of CC-BY-4.0 data spanning simulation, benchtop suturing, ex vivo tissue, animal work and real clinical procedures.
One contributor dominates. CMR Surgical, the Cambridge, England maker of the Versius system, supplied close to 500 hours of anonymized video and robotic telemetry, the majority of the surgical corpus by duration.
"Surgical robotics generates a rich understanding of how procedures are performed," said Chris Fryer, CMR Surgical's chief technology officer, in a statement issued the day after the launch. "By contributing real-world surgical data to collaborative initiatives like Open-H, we are helping build the foundations for the next generation of intelligent surgical systems. Because Versius is the most software-driven robot on the market, we are well-placed to share our data with the wider ecosystem."
What the VLA actually does
GR00T-H is a 3-billion-parameter model that takes camera images and a plain-language task description and emits action tokens — continuous motor commands for a specific robot. The backbone is Cosmos Reason 2 2B; a diffusion transformer head denoises the action trajectory using flow matching. The interesting engineering is the cross-embodiment plumbing: a learnable MLP per robot maps that machine's kinematics into a shared normalized action space, and at inference the model throws away proprioceptive input entirely, substituting a learned per-hardware bias term. For cable-driven arms with mechanical hysteresis, trusting the camera over the joint encoders works better.
The benchmark numbers are modest and stated as such. On SutureBot, GR00T-H completed a full end-to-end suture in 25% of 20 trials — against zero for ACT and for the prior GR00T-N1.6 baseline. On out-of-distribution wound configurations and lighting, it averaged 54% sub-task success versus 30% for the earlier foundation model and 5% for ACT. A quarter of attempts is not a clinical standard. It is the first time any foundation model has finished the sequence at all.
Cosmos-H addresses the same bottleneck from the other side. The surgical simulator was fine-tuned from Cosmos Predict 2.5 2B on 64 A100s for roughly 10,000 GPU-hours, across nine robot embodiments and 32 datasets. Nvidia reports 600 policy rollouts taking 40 minutes in simulation against two days on a real benchtop rig. Rheo sits a level up: a digital twin of clinical workflows, device interactions, human movement and hospital logistics, so developers train the building rather than the OR table.
Why it matters
This is the platform play, run again. Nvidia does not make a surgical robot and shows no sign of wanting to. It is trying to own the layer every surgical robot maker will build on — dataset, pretrained policy, simulator, digital twin — and to make that layer permissively licensed enough that declining it is expensive. The June follow-up made the intent explicit: GR00T-H-N1.7 shipped on June 17 under the Nvidia Open Model License, post-trained on 601 hours of real surgical tasks across 58 datasets and seven platforms, and legal to deploy commercially. The research license was the demo; the commercial license is the business.
The structural consequence is that data hoarding stops being a moat. If the shared base exists, advantage moves to fine-tuning and regulatory throughput — who validates fastest, not who collected most.
"The next generation of surgical robotics will be powered by data, simulation and AI working together," said David Niewolny, Nvidia's head of business development for healthcare and medical technology. The aspirational phrase in the company's literature — a "surgical robotics ChatGPT moment" — refers to a future Open-H version with annotated reasoning traces, not to anything shipped.
Nothing here is cleared for clinical use. GR00T-H-N1.7's model card says so directly. FDA guidance still holds that robotically assisted surgical devices cannot operate without direct human control, and a 2024 systematic review found 86% of cleared surgical robots sitting at Level 1 assistance. The 778-hour corpus is enormous for this field and trivial against the 20,000-plus hours of human egocentric video used to pretrain the base GR00T model.
What to watch
Whether Rheo produces deployments or demos. Proximie said in April its Smart OR platform would be built on the blueprint; Apian is building NHS hospital twins on it; PeritasAI is using it for OR humanoids, with Zimmer Biomet evaluating orchestration. Watch CMR's regulatory calendar too — a gynecology 510(k) went in this April, and the July SRS demo of Cosmos-H-Dreams predicting the surgical field under proposed actions is the first sign of this stack pointing at a cleared device. "Today, surgical robots are tools in the operating room," Fryer said then. The gap between that sentence and the next one is measured in trials, not releases. And watch Open-H v2: if the reasoning-trace dataset lands with the same contributor base, the moat question gets answered for the decade.
“Because Versius is the most software-driven robot on the market, we are well-placed to share our data with the wider ecosystem.”— Chris Fryer, Chief Technology Officer, CMR Surgical