A new, openly available collection of robotics data is setting the stage for a more hands-on generation of medical AI. For years, healthcare AI has excelled at perception—reading scans and identifying anatomy—but the physical act of intervention has lagged behind. A coalition of 35 institutions, led by figures from Johns Hopkins, TU Munich, and NVIDIA, has released Open-H-Embodiment to address that gap.

The dataset offers 778 hours of licensed surgical, ultrasound, and colonoscopy robotics data, recorded from both commercial systems and research platforms. It pairs real clinical procedures with simulation and benchtop exercises, creating a resource for training AI that doesn't just see, but acts.

Alongside the data, the group released two open-source models. GR00T-H is a policy model trained to control surgical robots, using novel techniques to manage the quirks of different hardware. In tests, a prototype successfully completed an end-to-end suturing task. A second model, Cosmos-H-Surgical-Simulator, acts as a predictive world model, generating realistic surgical video from robot motions to speed up training and bridge the gap between simulation and reality.

The initiative's organizers describe the next goal as moving toward 'reasoning-capable autonomy'—systems that can explain decisions and adapt during long procedures. They are calling for broader community involvement to expand the dataset with annotated task traces, aiming to build a foundation for the next leap in robotic assistance.

Source: Hugging Face Blog