Perception Uncertainty and Collective Robustness in AI-Driven Swarm UAV Systems

This paper examines risk formation and propagation in AI-driven swarm-based unmanned aerial vehicle (UAV) systems operating in complex environments. In distributed multiagent systems, perception errors made by individual agents can propagate through collective interaction and amplify at the swarm level. The lack of a structured approach to assessing such risks increases uncertainty during the preparation, training, and operational deployment of AI-based Geospatial Perception Models (AGPM). A methodological framework is introduced to classify risks across the AGPM lifecycle, including knowledge selection, knowledge analysis, model training, and operational stages. The framework considers both human and technological factors and applies the Analytic Hierarchy Process (AHP) to assess their relative significance within a hierarchical multiagent setting. Particular attention is paid to cascading error propagation and its impact on collective swarm behavior. Experiments with contemporary general-purpose multimodal AI systems were conducted to evaluate their ability to interpret graphical representations of complex geospatial scenes relevant to autonomous navigation. The results suggest that while semantic recognition remains stable, spatial-geometric reasoning and formal uncertainty estimation are limited. In swarm-based systems, these limitations may contribute to the amplification of local perception errors. Overall, the analysis indicates that technological risks during operation, especially cascading error propagation among agents, have the strongest influence on the overall risk profile. The framework is intended to clarify how vulnerabilities emerging at different lifecycle stages may affect collective behavior in AI-driven swarm systems.