Safe Variable-Force Control for Robotic Ultrasound Scanning via ℓ2-Regularized Fuzzy Broad Offline Reinforcement Learning

Abstract

Robotic ultrasound scanning (RUSS) is a promising technology for standardizing examinations, improving repeatability, and alleviating the workload of sonographers. A key prerequisite for its safe clinical use is reliable control of the contact force between the probe and the patient’s skin. In practice, this force should vary along the scan: it must be reduced over symptomatic or painful regions to avoid discomfort and potential tissue damage, while remaining sufficiently high elsewhere to ensure good acoustic coupling and image quality. Patient-specific soft-tissue properties and unstructured examination environments make it difficult to tune conventional (adaptive) impedance controllers, often leading to force overshoot or poor tracking performance. To address this challenge, we propose an $\ell _{2}$ -regularized fuzzy broad offline reinforcement learning framework ( $\ell _{2}$ -FBORL) that learns variable force control policies from limited, suboptimal demonstrations generated by stabilizing controllers. Building on $\ell _{2}$ -FBORL, we formulate variable force tracking in RUSS as a Markov decision process and develop a Fuzzy Learning-based Adaptive Impedance Control (FLAIC) scheme that adaptively adjusts the impedance adaptation rate to suppress overshoot while maintaining fast convergence. The proposed approach is evaluated in high-fidelity simulation, phantom experiments, and in vivo upper-limb venous and abdominal scanning under respiratory motion. Across all scenarios, FLAIC reduces force overshoot and improves tracking accuracy compared with conventional impedance and adaptive impedance controllers, without patient-specific retuning, indicating its potential to enhance the safety, robustness, and standardization of robotic ultrasound procedures.