NVIDIA has announced Cosmos-H-Dreams, a real-time, action-conditioned generative simulation platform designed to accelerate the development of surgical robotics. As robotic surgery advances toward complex vision-language-action policies, training and evaluation remain bottlenecked by the high cost of physical platforms, slow experimentation cycles, and the risk of damaging biological tissue or instruments. Traditional simulators struggle to accurately render deformable tissues, specular surfaces, and fine mechanical interactions. Cosmos-H-Dreams addresses these challenges by transforming a previously offline surgical world model into an interactive, streaming environment that operates within a closed-loop control system. Built upon the foundational capabilities of NVIDIA’s earlier Cosmos-H-Surgical-Simulator, the new platform utilizes a teacher-to-student distillation pipeline to preserve critical surgical dynamics while drastically reducing computational overhead. The model undergoes self-forcing distillation, training a causal student network to imitate autoregressive rollouts while conditioning on its own previously generated frames. This approach enables the student to operate with as few as two denoising steps per latent frame. To support high-frequency interaction, NVIDIA integrated the model with FlashDreams, an optimized inference library featuring streaming key-value caches and CUDA Graph capture. Deployed on a single NVIDIA RTX PRO 6000 GPU, the system achieves approximately 160 frames per second, a substantial leap from the ten frames per second typical of standard diffusion-based surgical simulators. The platform is engineered for broad integration and extensibility. Initial deployments demonstrate compatibility with the da Vinci Research Kit tabletop suturing framework and have been successfully piloted alongside CMR Surgical’s Versius surgeon controller. NVIDIA has also published a comprehensive training recipe, allowing researchers to fine-tune the distilled student model for custom embodiments and specialized surgical datasets. User interaction is facilitated through multiple interfaces, including WebRTC-enabled browser clients, Meta Quest controllers mapped via WebXR, and direct policy-to-policy closed-loop connections. Cosmos-H-Dreams establishes a new benchmark for evaluating and training surgical automation. By generating rare failure scenarios on demand and maintaining structural consistency over extended horizons, the simulator supports rigorous closed-loop benchmarks for tool-tip accuracy, gripper fidelity, and counterfactual action diversity. This capability significantly reduces reliance on physical hardware for reinforcement learning and imitation training. Looking ahead, the architecture is positioned to support latency-aware telesurgery, interactive procedural rehearsal, and intraoperative planning. NVIDIA explicitly notes that Cosmos-H-Dreams is strictly a research and development platform, not a clinical diagnostic tool or physical robot controller. As temporal stability and hardware efficiency improve, the system aims to unify surgeon education, synthetic data generation, and policy validation within a single, scalable Physical AI ecosystem.