Hardware-Agnostic Imitation Learning Method for Autonomous Ultrasound Scanning Addressing Physical Deployment Discrepancies.

To achieve autonomous ultrasound scanning skill transfer across different physical equipment instances and address the limitations of traditional imitation learning methods-which struggle with cross-instance generalization due to their reliance on specific manipulator parameters-this study proposes a physical-parameter-decoupled imitation learning method based on waypoint representation. This approach utilizes a greedy algorithm to automatically extract key nodes within the task space from expert demonstration trajectories, constructing a trajectory representation decoupled from low-level kinematic parameters and base calibration errors. Simultaneously, a velocity-aware adaptive error precision adjustment mechanism is introduced to dynamically modulate waypoint extraction thresholds, simulating the speed-accuracy strategies employed by sonographers across different scanning phases. Cross-validation across two mainstream generative architectures-Action Chunking Transformer (ACT) and Diffusion Policy-on an offline dataset confirms the plug-and-play capability of waypoint representation in suppressing long-horizon error accumulation, with both architectures achieving significant reductions in prediction errors. For physical deployment, a complete ACT-waypoint system featuring low-level triple safety redundancy was validated. In kidney long-axis standard plane scanning tasks, the system achieved a 92% success rate on the source domain manipulator and maintained an 84% success rate on the target deployment manipulator, despite incompatible low-level kinematic parameters and base coordinates. Force control accuracy remained stable around the target value of 12 N. The results demonstrate that the proposed method effectively overcomes base coordinate and D-H parameter discrepancies to achieve cross-instance zero-shot skill transfer, significantly enhancing the adaptability across physical instances and the scanning success rate of imitation learning models.