NVIDIA open-sources a surgical-robot simulator that runs 16x faster than the model it was distilled from

NVIDIA published Cosmos-H-Dreams on 27 July 2026 — an action-conditioned generative simulator that takes a surgical scene plus a sequence of robot actions and generates the future video, so a policy can be trained and evaluated without theatre time. NVIDIA reports roughly 160 frames per second on a single RTX PRO 6000 through its FlashDreams inference library, against roughly ten for the larger Cosmos-H-Surgical-Simulator it was distilled from, using as few as two denoising steps per latent frame. Checkpoints and code are on Hugging Face and GitHub; it accepts keyboard and VR-controller input or a learned policy, across the dVRK tabletop suturing setup and the Versius platform, with CMR Surgical and Cambridge Consultants named as collaborators.

What it means Surgical-robot policy work has been gated on scarce hardware time; a faster-than-real-time simulator with published weights moves that loop onto one workstation GPU.