At its GTC conference on March 16, 2026, Nvidia unveiled a set of open datasets and models aimed squarely at a problem that has slowed surgical robotics for years: there simply isn’t enough recorded, standardized data of real operations for AI systems to learn from. The centerpiece is Open-H, a dataset that has grown to roughly 778 hours of surgical video and robotic motion data pooled from 35 contributing organizations, including surgical robot makers CMR Surgical, Rob Surgical, and Tuodao, along with academic research platforms like Johns Hopkins’ dVRK and the Franka robotic arm.
Why Surgical Robots Need a Data Set Like This
Training an AI model to guide or assist a robotic arm during surgery requires far more than images; it requires synchronized video, instrument position data, and outcome context, all recorded at consistent resolution and frame rates across many different procedures and hospitals. Historically, that kind of data has stayed locked inside individual device makers and hospital systems, each using its own formats and rarely sharing footage outside its own research programs. Nvidia’s pitch with Open-H is to standardize that data, in the LeRobot v2.1 format at a minimum of 20 frames per second and 480p resolution, so that models trained on one robotic platform can be evaluated or adapted for another, a capability known as cross-embodiment learning.
The Models Built on Top of the Data
Nvidia paired the dataset with GR00T-H, a 3-billion-parameter vision-language-action model trained on a 601.5-hour subset of Open-H drawn from 58 individual datasets. GR00T-H is designed to take in a text description of a clinical task and output the motion commands a robotic system would need to attempt it, and it currently supports several real surgical robot platforms, including CMR Surgical’s Versius, the dVRK and dVRK-Si research platforms, UR5 arms, Rob Surgical’s Bitrack, Tuodao’s MA2000, and KUKA hardware. Nvidia has been explicit that GR00T-H is intended for research and development only, not for use in an actual operating room on a patient.
Simulating Surgery Before Any Robot Touches a Patient
Alongside the data and policy model, Nvidia released Cosmos-H-Surgical-Simulator, a world model built on its Cosmos-Predict2.5 architecture that can generate physically plausible video of a surgical scene evolving in response to a robot’s planned actions. In practical terms, that lets researchers test how a robotic policy would behave across hundreds of simulated variations of a procedure without needing a real patient, cadaver, or benchtop setup for each one. Nvidia says the system, trained using roughly 10,000 GPU-hours across 64 A100 chips, can generate 600 simulated rollouts of a procedure in about 40 minutes, compared with roughly two days to run the equivalent trials using physical benchtop testing methods.
An Early Industry Partner
CMR Surgical, the British maker of the Versius surgical robot already used in operating rooms across more than a dozen countries, has contributed close to 500 hours of surgical video to Open-H and says it is using Cosmos-H’s simulation capabilities to train and validate robotic intelligence for Versius before any of that intelligence reaches an actual clinical deployment. That sequencing, simulate and validate extensively before going anywhere near a live procedure, reflects the far higher safety bar surgical robotics faces compared with other robotics applications Nvidia has targeted, such as warehouse or manufacturing automation.
Skepticism and the Distance to the Operating Room
Robotics researchers outside Nvidia have welcomed the standardized dataset as a genuine gap-filler but have cautioned that synthetic, simulation-generated training data still faces a known problem in robotics called the “sim-to-real gap,” where a policy that performs well in simulation underperforms when it meets the variability of an actual human body and a live surgical team. Because GR00T-H is explicitly restricted to research use, none of what Nvidia announced translates into an autonomous surgical robot reaching patients any time soon; the near-term impact is on how surgical robot makers develop and test their own proprietary, FDA-regulated systems, not on what a patient encounters on the operating table today.
What to Watch Next
Nvidia and its partners have signaled that Open-H will keep growing as more device makers and hospitals contribute footage, and the real signal of progress will be whether companies like CMR Surgical, Rob Surgical, and others start citing Open-H-trained or Cosmos-H-validated components in their own FDA submissions over the next one to two years. Until that happens, the project remains an infrastructure play, an attempt to give the surgical robotics field the kind of large, shared, standardized dataset that fields like autonomous driving and general-purpose robotics already have.
