This project builds on previous research that developed AI-based navigation systems for uncrewed aerial vehicles (UAVs) in simulated environments. While earlier work demonstrated that reinforcement learning can enable efficient and safe navigation under challenging conditions. The main objective of this study is to enhance the interpretability of reinforcement learning-based navigation models. A lightweight module is developed to provide real-time decision logging and visual explanations of UAV flight paths. These tools allow us to better understand how the model responds to changing environmental conditions, making it easier to analyze its behavior during operation.

A key focus of the project is to evaluate how trained models generalize beyond simulation. Real-world environments often include uncertainties such as ambiguous terrain, occlusion, and degraded GPS signals, which are difficult to fully replicate in simulation. By conducting simulations with real-world data collected with a compact UAV in a controlled test area, this project use interpretability tools to examine system performance under these conditions.

This research emphasizes improving transparency and traceability. The findings are expected to support more reliable deployment of AI-driven UAV systems and contribute to the development of responsible autonomous technologies. Overall, this project extends previous work by addressing a critical gap between simulation performance and real-world applicability.