Closing the Sim-to-Real Gap: An End-to-End Robotic Ultrasound System Leveraging In Vivo Reinforcement Learning and 3D-Prior Guided Hybrid Control

Abstract

Abdominal ultrasound is a crucial first-line diagnostic tool, yet its efficacy is inherently constrained by a strong dependency on operator skill, leading to significant inter-operator variability and limiting its widespread adoption. While robotic ultrasound systems with artificial intelligence have emerged to mitigate this, current approaches face critical limitations: they often rely on simulated or phantom data for training, which hampers generalizability, and employ control strategies that lack the flexibility to adapt to complex human anatomy. To address these challenges, this paper introduces a novel, end-to-end robotic system for autonomous abdominal scanning. Our methodology is systematic: it begins with coarse scan planning via RGB and 3D point cloud-based localization of key anatomical landmarks (e.g., xiphoid process). The core of our approach involves training a reinforcement learning scanning policy directly on live human volunteers, enabling the development of a strategy that is robust to anatomical diversity. This high-level policy is executed by a sophisticated hybrid force-position controller, enhanced with real-time normal vector calibration and 3D point cloud-based respiratory phase detection to dynamically adjust contact force, particularly during inspiration. This innovation proves critical for improving scan quality in subjects with higher Body Mass Index (BMI). Extensive validation against expert sonographers demonstrates that our system achieves high precision in standard plane recognition and superior scanning efficiency. Furthermore, to holistically assess performance where standard planes are difficult to acquire, we introduce a 3D reconstruction-based coverage metric. Results confirm that our system delivers strong adaptability and significantly improved consistency, marking a substantial step toward clinically viable, operator-independent ultrasound robotics.