Nvidia's Cosmos-H-Dreams simulates surgical robots in real time on a single GPU

Nvidia published Cosmos-H-Dreams on Hugging Face on July 27, 2026, a real-time, action-conditioned generative simulator for surgical robotics distilled from the company's earlier Cosmos-H-Surgical-Simulator world model. Where the original teacher model runs at roughly 10 frames per second, the new causal student model reaches about 160 frames per second on a single Nvidia RTX PRO 6000 GPU, using as few as two denoising steps per frame, fast enough to support real-time, closed-loop control by either a human operator or an automated policy. Nvidia trained the system in three stages: first fine-tuning a surgical teacher model on data from Johns Hopkins' dVRK surgical robot, including realistic failure cases like dropped needles and missed suturing throws, then warming up a faster student model on the teacher's precomputed outputs, and finally letting the student generate its own training data while the frozen teacher corrects its mistakes. The release includes a pretrained checkpoint for tabletop suturing tasks and was demonstrated working with CMR Surgical's Versius robotic platform through a partnership with Cambridge Consultants, built on the broader Open-H-Embodiment surgical dataset Nvidia and academic partners released earlier this year. Nvidia was explicit that Cosmos-H-Dreams is “a research and development platform, not a diagnostic system, a replacement for intraoperative imaging, or a controller for a physical surgical robot,” positioning it instead for surgical rehearsal, procedure planning, and training rather than live operations. The release fits a broader push by Nvidia this year to bring its Cosmos world-model line into specialized physical domains beyond autonomous driving and warehouse robotics.