Tinyml-Empowered Climbing Robots for Real-Time Structural Defect Detection: Case Studies on Bolt and Tile Inspection

Structural inspection of critical infrastructure presents significant safety challenges when performed manually, particularly for bolt connections in steel structures and exterior wall tiles. This paper presents an innovative approach integrating Tiny Machine Learning (TinyML) with climbing robots to enable autonomous, real-time structural defect detection in resourceconstrained environments. Two complementary case studies are examined: a magnetic climbing robot using the FOMO algorithm for visual bolt defect inspection in steel structures and a vacuumbased climbing robot employing CNN-based acoustic analysis for tile defect detection on building exteriors. The magnetic robot achieved 82 % accuracy with an inference time of just 86 ms while requiring minimal memory resources <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{(} \mathbf{0. 0 7 2 M B}$</tex> flash, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2 3 9. 1 K B}$</tex> RAM). The vacuum-based robot demonstrated 74.62 % accuracy in distinguishing between defective and intact tiles through acoustic signature analysis. Both implementations showcase TinyML's versatility across different sensing modalities and structural inspection domains. This approach enhances inspection accuracy and consistency and significantly improves safety by eliminating human exposure to hazardous environments. The methods presented demonstrate TinyML's broader applicability for autonomous inspection tasks in resource-constrained robotic systems.