Robot-assisted ultrasound imaging towards embodied intelligence: a survey

Robot-assisted ultrasound imaging has emerged as an important direction for improving the consistency, accessibility, and safety of ultrasound examinations. Unlike conventional image interpretation tasks, ultrasound image quality is determined by the closed-loop interaction among probe motion, contact force, tissue deformation, acoustic coupling, and clinical intent, making robot-assisted ultrasound imaging a representative scenario for embodied intelligence. In this survey, embodied intelligence refers to the coordination of sensing, actuation, and adaptive decision-making within a physical perception-action loop that links probe manipulation to imaging outcomes. Rather than reviewing artificial intelligence algorithms or robotic platforms in isolation, this paper analyzes how sonographer-like scanning intelligence is progressively externalized into robotic morphology, sensing architectures, control strategies, simulation environments, and learning-based autonomy. We first summarize clinical application scenarios and their task-specific embodied requirements. We then review foundational technologies, including mechanical design, actuation mechanisms, ultrasound and external sensing, and force-position interaction control. We emphasize sensor-actuator integration, contact-force measurement, mechanical compliance, and engineering trade-offs that influence scanning performance. Ultrasound imaging and scanning simulation are discussed as infrastructures for modeling action-dependent image formation and supporting policy learning. Recent advances in reinforcement learning, imitation learning, and methods based on large language models are examined from the perspective of embodied perception-action coupling. Finally, we discuss future challenges in patient-safe embodiment, calibrated multimodal perception, measurable sim-to-real transfer, clinically supervised autonomy, and translational validation.