
Tech • AI • Robotics
NVIDIA is expanding its healthcare push around physical AI, combining open robotics data, world models, simulation tools and deployment hardware to accelerate surgical robots, hospital automation and AI-driven medical imaging.
Healthcare has moved from imaging and image reconstruction to deep learning, ambient clinical AI and now physical AI, where systems can perceive and act in real environments. The need is large: roughly 160,000 hospitals, 72,000 procedures and 8 million medical devices operate in settings where care is ultimately delivered physically, not just digitally.
The current strategy rests on three technical layers: world foundation models such as Cosmos, robotic policy models such as Groot and Groot H, and high-fidelity physics simulation. Together, these tools aim to let healthcare robots understand clinical scenes, reason about tasks and practice actions before being used around patients.
A major step is Open AHE, described as one of the largest open healthcare robotics data releases so far. It includes more than 35 partners, data across 11 embodiments, and over 750 hours of healthcare robotics footage and signals, with contributors including CMR Surgical and Moon Surgical. The dataset is intended to ground training in real clinical physics before synthetic data generation expands coverage.
Developers argued that real-world data alone cannot cover every anatomy, device setup or edge case, especially in surgery. The proposed answer is a data flywheel that starts with real data, then continuously post-trains models with large amounts of synthetic data from both classical simulation and neural simulation, improving robustness across rare scenarios.
Groot H, released as open source on GitHub with weights on Hugging Face, adapts general robot action models to surgical settings. The architecture combines a Cosmos reasoning engine that interprets vision and language with a diffusion transformer that converts robot state and semantic understanding into action policies. Developers said generic vision-action systems do not generalize well to surgery without domain-specific post-training.
Cosmos H adds healthcare-specific world modeling for controllable synthetic data generation, surgical state prediction and learnable simulation. It can start from digital twins and anatomy grounded in CT and MRI data, then generate varied laparoscopic scenes and predict different action states, giving researchers a faster way to test policies than repeated lab validation on physical robots.
These assets are being distributed through Isaac for Healthcare, a domain-specific extension of NVIDIA Isaac. The package includes sensor simulation libraries, synthetic data pipelines, policy models, developer blueprints and end-to-end workflows aimed at surgical robotics, medical devices and hospital automation.
Version 0.5 of Isaac for Healthcare adds sensor simulation libraries for ultrasound and X-ray. The ultrasound tools model tissue acoustic properties and probe configuration to generate realistic scans, while the X-ray release simulates the pipeline from CT-derived volumes through ray propagation and detector response, supporting synthetic sensor data generation at scale.
Companies including CMR Surgical, Johnson & Johnson MedTech, Moon Surgical, Medbot, Exar Labs and Lem Surgical are using parts of the stack for digital twins, synthetic data and subtask automation. Developers said 2026 could mark a broader deployment phase for healthcare robots, though fully autonomous surgery remains a long-term goal, with nearer progress expected in limited subtask automation.
Holoscan 4.0 and IGX Thor were positioned as the deployment layer for physical AI. Holoscan 4.0 adds EtherCAT, interoperability with tools such as ROS 2 and GPU-resident graphs for lower latency, while IGX Thor offers about eight times more compute for edge AI workloads expected to run future vision-language and vision-language-action models.
NVIDIA also introduced medical imaging models that move AI closer to sensor physics. New raw-to-insight systems for ultrasound and MRI aim to work directly from raw acquisition data toward reconstruction and interpretation, while NV Generate for CT and MR is a 3D diffusion model designed to create synthetic anatomical data for training, simulation and digital twin workflows rather than clinical diagnosis.
The broader goal is to make healthcare robotics and medical AI faster to train, cheaper to validate and easier to deploy. The main test will be whether open data, synthetic simulation and edge deployment tools can close the gap between lab performance and reliable use in hospitals.
Explain this