Robotic ultrasound could standardize image acquisition and reduce operator workload, but training such robots requires realistic simulation environments. Existing simulators often rely on robot-tracked poses and latent-space prediction, and they don't directly generate the visual observations needed for interactive control. One reason is that each ultrasound frame is only a cross-section of unseen 3D anatomy, so even small probe movements can produce large changes in the image.
The authors propose SonoWorld, a causal ultrasound world model. It is trained in three stages: first, a latent-action-conditioned video diffusion model is pretrained on roughly 6.9 million unlabeled multi-organ ultrasound frames; second, the learned video dynamics are grounded to metric probe poses using a small dataset of EM-tracked freehand thyroid scans; third, distribution matching distillation is applied to allow causal, long-horizon autoregressive generation with selective caching of eviction tokens at inference.
The team evaluated controllability using pose-following fidelity, round-trip consistency, and axis-specific control under synthetic actions on held-out tracked scans, and also ran a user study. In a model-based action discrimination task, SonoWorld correctly ranked target-directed probe motion over an equal-magnitude opposite motion in up to 74.8% of trials. The authors say this is the first causal, pose-controllable ultrasound world model learned primarily from freehand clinical videos, offering a data-driven interactive environment for robotic ultrasound data acquisition and policy evaluation.