Navigation and Localization for Railway Inspection Drone in GPS-denied Environments: An Investigative Study of Modular SLAM Baselines and End-to-End Learning-based Approaches

Abstract

Autonomous navigation in GPS-denied environments remains a critical challenge for unmanned aerial vehicles performing infrastructure inspection. This thesis investigates the feasibility of learning-based navigation for railway-inspection drones by first constructing and analyzing a state-of-the-art modular baseline and then evaluating emerging end-to-end paradigms. A high-performance navigation system combining FasterLIO SLAM with Fast-Planner trajectory generation is implemented in high-fidelity simulation and used as an analytical baseline. While the system successfully navigates dense forest environments, controlled experiments reveal three structural failure modes—SLAM localization drift, flight-controller tracking limitations, and planner-induced trajectory constraints—highlighting deeper challenges such as cumulative error propagation and real-time sensor–compute bottlenecks. Building on these insights, the thesis conducts an experimentally grounded feasibility study of three dominant end-to-end learning directions: predictive worldmodel architectures, self-supervised representation learning pipelines, and visionreinforcement- learning approaches. By implementing prototype models and stresstesting their stability and data requirements, the study identifies several infeasible or unstable directions—such as feature-forecasting models and self-distillation objectives— and reveals simulator limitations that currently block scalable vision-RL for UAVs. Rather than delivering a complete end-to-end navigation system, this work provides a systematic evaluation of the landscape, clarifies the fundamental obstacles facing learning-based navigation in GPS-denied environments, and establishes concrete design requirements and a research roadmap for future PhD-level research.