Confidential-weighted cooperative merging of observations for safe navigation of UAV swarms

The subject of the research is the development and justification of the methodology for trust-weighted cooperative observation fusion in the task of safe navigation of swarms of unmanned aerial vehicles (UAVs). The object of the study is multi-agent systems operating under the constraints of limited computational resources, instability of communication channels, and heterogeneity of sensor data. The author thoroughly examines aspects such as the asynchrony of information exchange, uncertainty and uncalibrated outputs of perception algorithms, and the impact of overconfident yet inaccurate data sources on the overall perception picture. Special attention is given to creating a statistically reproducible environment for validating algorithms, enabling comparisons of the effectiveness of independent perception, naive cooperation, and trust-weighted fusion. The analysis focuses on the architecture of CONTUR-OD (Cooperative Explainable Neural Network Trajectory Node of the Swarm with Trust Assessment), which provides explainability of decisions and allows flexible modeling of the entire cycle of swarm operation: from perception and communication to trajectory planning and collision avoidance. The research methodology is based on statistical approximation of object detectors, trust-weighted observation fusion, and reactive trajectory planning based on potential fields and barrier functions. Experiments were conducted in a simulation with controlled parameters of sensors, communication, and threat dynamics. The main conclusions of the study confirm that cooperative perception significantly enhances swarm safety compared to independent strategies, while trust-weighted fusion further reduces collision frequency and improves trajectory consistency relative to naive averaging. A significant contribution of the author to the research topic is the proposal of the CONTUR-OD architecture, which enables scalable and reproducible experiments without the excessive computational costs typical of heavy neural network models. The novelty of the research lies in the formalization of the "perception – communication – fusion – planning – movement" cycle as a modular pipeline with calibrated measures of confidence and a transparent causal relationship between the quality of perception and navigation safety. The results showed a reduction in collision frequency by more than 50% and an improvement in minimum inter-object distances with a moderate increase in mission time, confirming the effectiveness of the trust-weighted approach and its promise for real scenarios in swarm management of UAVs.