EmbodiedUS-FS: Fast-Slow Intelligence for Ultrasound Robotics

Robotic ultrasound scanning requires balancing two distinct objectives: high-level clinical reasoning (understanding physician intent and procedural protocols) and low-level reactive control (maintaining image quality despite patient motion or target drift). The authors propose a "Fast-Slow" hierarchical architecture to decouple these tasks. The Slow Brain handles deliberative planning, utilizing knowledge retrieval from clinical handbooks and APIs to construct a structured task graph. The Fast Brain operates at a high frequency, performing closed-loop control by fusing multimodal feedback—such as ultrasound images, robot pose, and contact force—to refine actions within the constraints set by the Slow Brain.

To ensure safety and reliability, the system incorporates three critical components:

• Structured Plan Verification: A module that checks the Slow Brain's generated plans against API legality, parameter validity, and workflow consistency before execution.
• Safety Shield: A constraint layer that restricts risky actions and monitors for safety-bound violations during real-time operation.
• Hierarchical Escalation Policy: A mechanism that triggers replanning or requests human confirmation if the Fast Brain encounters persistent failures, high uncertainty, or safety violations that it cannot resolve locally.

Existing robotic ultrasound systems often struggle to reconcile the semantic complexity of clinical instructions with the dynamic requirements of physical scanning. By separating these into a deliberative, knowledge-grounded planner and a reactive, feedback-driven controller, this framework provides a more interpretable and robust approach to autonomous medical assistance. The integration of safety-aware escalation ensures that the robot remains controllable even when faced with the unpredictable nature of clinical environments.