NVIDIA Releases an Open AI Stack for Surgical Robotics

NVIDIA Developer

Building an autonomous surgical robot requires more than advanced hardware. It also requires large datasets, realistic simulation environments, and AI models capable of learning physical tasks.

At GTC 2026, NVIDIA introduced an open platform designed to provide those building blocks. The release includes a large surgical dataset, generative models for synthetic data, a vision-language-action model for robotic control, and a simulation framework for healthcare environments.

Several surgical robotics companies, including CMR Surgical, Johnson & Johnson MedTech, Moon Surgical, Rob Surgical, PeritasAI, and Proximie, have already begun integrating the platform into their development pipelines.

Open-H

One of the most significant announcements is Open-H, the largest open dataset created specifically for healthcare robotics.

Built with 35 collaborators, the dataset contains 776 hours of surgical video collected across 11 robotic systems and four surgical indications. The goal is to provide developers with a common dataset for training and evaluating AI models that can generalize across different procedures and robotic platforms.

Large, diverse datasets have been fundamental to progress in computer vision and large language models. NVIDIA is positioning Open-H to play a similar role in surgical robotics.

Cosmos-H

Training robotic AI also requires data that does not yet exist.

Cosmos-H generates realistic surgical videos from text prompts, reference images, existing videos, or robot kinematics. Developers can use these synthetic procedures to expand training datasets, simulate uncommon clinical scenarios, and evaluate robotic policies before deploying them in physical systems.

As robotic AI models become larger, synthetic data is expected to become an increasingly important part of development.

GR00T-H

NVIDIA also introduced GR00T-H, a vision-language-action model trained on Open-H.

The model interprets clinical instructions and converts them into motion commands for robotic systems. Rather than focusing only on recognizing surgical scenes, it is designed to connect perception with physical action, allowing developers to train and evaluate robots that perform complex tasks in healthcare environments.

Rheo

The final component is Rheo, a simulation framework that creates physically accurate digital hospitals.

Developers can simulate operating rooms, clinical workflows, medical devices, hospital logistics, and interactions between healthcare professionals. These virtual environments allow robotic systems to be trained and tested extensively before entering clinical settings.

Early Adopters

Several companies are already applying these technologies.

CMR Surgical contributed nearly 500 hours of surgical video to Open-H and is using Cosmos-H to generate synthetic surgical data and evaluate robotic policies.

Johnson & Johnson MedTech is using NVIDIA's simulation technology to generate training data for the MONARCH Platform for Urology.

PeritasAI is combining Isaac for Healthcare and Rheo to develop embodied AI systems that coordinate robots, surgical instruments, implants, and operating room workflows.

Proximie is using Cosmos-H to train multimodal vision-language models that analyze operating room images and intraoperative video, supporting AI agents that assist surgical teams during procedures.

Looking Ahead

NVIDIA is not introducing another surgical robot. It is providing the infrastructure that allows researchers and medical device companies to build them more efficiently.

Open-H, Cosmos-H, and GR00T-H are available through GitHub and Hugging Face, giving developers access to tools that were previously unavailable at this scale.

Whether these models ultimately accelerate the development of autonomous surgery will depend on how widely they are adopted. What is clear is that surgical robotics now has an open AI foundation that did not exist before.

Source: NVIDIA Developer