Nvidia Builds an Open Foundation for Surgical Robots With Open-H Video and Cosmos-H Synthetic Worlds

Nvidia Builds an Open Foundation for Surgical Robots With Open-H Video and Cosmos-H Synthetic Worlds

The hardest thing about teaching a robot to operate on a person is that you cannot afford to let it practice on one. Anatomy varies from patient to patient, instruments bend and slip against living tissue, and the rare complications a system most needs to anticipate almost never arrive on cue. Nvidia's answer, now open-sourced and in the hands of the world's largest medical-device makers, is to manufacture that experience: pair hundreds of hours of real surgical video with a generative engine that can dream up the operations a robot has never seen.

That is the bet behind the company's physical-AI platform for healthcare robotics, anchored by three pieces released across 2026 — Open-H, a dataset; Cosmos-H, a synthetic-data engine; and GR00T-H, a vision-language-action model. On July 22, Nvidia extended the stack by open-sourcing what it calls the first GPU-accelerated Medical Physics Simulation framework inside its Isaac for Healthcare toolkit, and revealed that CMR Surgical, Johnson & Johnson MedTech, Medtronic, XCath and others are already building on it.

What Nvidia actually shipped

Open-H — formally the Open-H Embodiment dataset — is the raw material. Nvidia describes it as the world's largest open dataset for healthcare robotics, combining real surgical video, robotic telemetry and multimodal data released under a permissive CC-BY-4.0 license. A June research paper describing its full scope reported roughly 780 hours of paired data drawn from 119 datasets, 20 robot platforms and more than 50 institutions, spanning surgical robotics, ultrasound and colonoscopy autonomy. The bulk of the surgical footage came from a single contributor: Cambridge-based CMR Surgical, which donated close to 500 hours of anonymized data from its Versius Surgical Robotic System, covering procedures including cholecystectomy, prostatectomy, hernia repair and hysterectomy.

That corpus underpins Isaac GR00T-H, which Nvidia bills as the first open vision-language-action model for healthcare robotics — a policy model, trained on roughly 600 hours of Open-H data, that turns text commands into precise physical motion.

Cosmos-H is where scarcity turns into abundance. The Cosmos-H-Surgical-Simulator is a world foundation model, fine-tuned from Nvidia's Cosmos Predict 2.5 base, that generates physically plausible surgical video directly from a robot's kinematic actions — including the messy realities that break conventional simulators, such as soft tissue, reflections, blood and smoke. A companion capability, Cosmos-H-Dreams, does this in real time: given the current state of the surgical field and a proposed action, it predicts what the field will look like afterward. The efficiency gains are stark. Nvidia says evaluating 600 policy rollouts takes about 40 minutes in simulation versus roughly two days on a real benchtop, and GPU-native simulation running 8,192 environments in parallel cut one training job from more than five hours to under two minutes. In its reference workflows, the company says more than 93 percent of the data used for policy training was generated synthetically.

Why the mix matters

The pairing is the point. Real surgical video is irreplaceable — it encodes how experienced surgeons actually handle tissue — but it is expensive, uneven and heavily skewed toward routine cases. Synthetic data fills the long tail: the anatomical variations and edge-case failures that a safe system must handle but that rarely appear in a hospital's recordings. Grounding a generative simulator in a large real-video foundation is what keeps those invented scenarios physically honest rather than hallucinated.

Nvidia's David Niewolny, who leads business development for healthcare and medical technology, framed the combination as the industry's new engine. "The next generation of surgical robotics will be powered by data, simulation and AI working together," he said. "By responsibly contributing surgical data and training open models on NVIDIA's physical AI platform, medical technology leaders like CMR Surgical are accelerating a new generation of intelligent robotic systems that can assist surgeons, scale surgical expertise and ultimately expand access to high-quality care."

The open-source posture is more than goodwill. In healthcare, model transparency doubles as regulatory infrastructure: access to data, weights and reproducible results is what lets developers evaluate performance across anatomies, document limitations and build evidence for approval. CMR Surgical's chief technology officer, Chris Fryer, made the case bluntly. "Open source models allow us to build on shared knowledge, accelerating responsible innovation and, ultimately, gives us the potential to deliver more consistent care and better outcomes for patients worldwide," he said.

A familiar Nvidia playbook

Strategically, this is the same move Nvidia has run in autonomous driving and general robotics, retargeted at the operating room. Rather than build a surgical robot, it supplies the substrate — datasets, world models, a learning framework and, of course, the GPUs everything runs on — and lets Johnson & Johnson, Medtronic, CMR and startups such as XCath and Inner Logic build products on top. J&J MedTech is using the physics simulation and a Cosmos-based foundation model to build digital twins of its endoluminal MONARCH urology platform, modeling kidney-stone scenarios; Medtronic's structural-heart group is exploring simulated X-ray sensing for catheter navigation. Each becomes a customer for the underlying stack. With the surgical-robotics market that J&J alone is chasing valued at more than $50 billion, controlling the training layer for physical AI is a durable position.

What to watch

The open question is whether policies trained overwhelmingly on synthetic data generalize to real patients — and whether regulators accept in-silico evidence as a legitimate part of the approval file. CMR's SRS 2026 demonstration of Versius Plus, which showed a system predicting how a procedure might unfold, hints at where this leads: robots that anticipate rather than merely follow. Watch for the first clinical deployments built on GR00T-H, for FDA signals on synthetic validation data, and for whether Nvidia's rivals answer with open datasets of their own or cede the foundation to Santa Clara.