Robotic Automation of Nasopharyngeal Swab Collection With Non-Contact Posture Adaptation

Abstract

Accurate nasopharyngeal specimen collection requires defining an appropriate insertion path, which is challenging under patient-specific facial anatomy and limited observation conditions. To address these challenges, this study presents an automated robotic framework that focuses on defining a patient-specific insertion path prior to the insertion process. The system uses an RGB-D camera and a deep learning model, EarPointNet, to detect facial landmarks, estimate the ear canal location, and compute a patient-specific insertion direction toward the nostril based on guidelines from the U.S. Centers for Disease Control and Prevention. EarPointNet achieved angular errors between 2.33° and 3.68° on an optimized dataset. Repeatability was further evaluated through comparative experiments with non-expert manual operation using a medical phantom. Additional experiments examined path estimation performance under partially occluded facial conditions with varying face-hole sizes, maintaining a mean angular error of 2.62° with a face hole radius of 100 mm. Future work will focus on integrating force sensing, torque sensing and reactive safety control, as well as conducting clinician-involved validation toward medical certification.