Quadruped robot-enabled framework for intelligent sidewalk condition monitoring

Abstract

Accurate detection of sidewalk conditions is critical for enhancing the safety and efficiency of urban non-motorized transportation systems. Existing manual inspection methods, while effective for limited scenarios, fall short in terms of efficiency, coverage, and the detection of concealed defects. This paper introduces a sidewalk inspection framework that leverages a quadruped robot equipped with visual and vibration sensors to comprehensively address complex pavement scenes and hidden loosened bricks. For visual inspection, an improved YOLO11 model with a dynamic attention-weighted sampling mechanism enhances detection accuracy for minority-class defects, achieving a mAP of 0.883. For hidden defect identification, a Physics-Informed Neural Network with a Long Short-Term Memory network (PINN-LSTM) is proposed to fuse physical constraints with temporal patterns, achieving an overall accuracy of 0.9430. The value of the physics-informed approach is further substantiated by its performance in cross-domain generalization, few-shot adaptation, and robustness tests, confirming its potential for practical applications.