Examination of Machine Learning-Based Visual Detect-and-Avoid for Small UAS: Datasets, Limitations, and Safety Assurance

Abstract

Visual detect-and-avoid functions are emerging as critical enablers for integrating small unmanned aircraft systems into shared airspace, where weight, power, and cost limitations favor passive electro-optical sensors and embedded machine learning inference. However, the properties of current learning-based perception pipelines, the datasets that support them, and the means of assuring their behavior across operational envelopes remain only partially characterized in a form suitable for rigorous safety assessment. This paper examines machine learning-based visual detect-and-avoid for small unmanned aircraft systems from three complementary perspectives: sensing and operational context, dataset construction and representativeness, and safety assurance under distributional uncertainty and rare-event conditions. The analysis focuses on end-to-end characteristics of camera-based detect-and-avoid pipelines, including detection, localization, and track-level decision logic, emphasizing conditions such as low-contrast targets, small apparent size, cluttered backgrounds, complex lighting, and non-cooperative traffic. Particular attention is given to the role of synthetic and hybrid datasets, temporal fusion, and uncertainty quantification as mechanisms to reduce systematic blind spots. The discussion is framed in terms of traceable performance requirements, interpretable failure modes, and integration of learning-based components into higher-level safety arguments without presuming unrealistic generalization guarantees. By consolidating modeling constructs and identifying constraints imposed by sensing geometry, computation, and data curation, the paper outlines practical considerations and technical limitations that must be addressed for learning-based visual detect-and-avoid to be deployed as part of credible risk-based approval pathways in small unmanned aircraft system operations.