Exploring feasibility of reinforcement learning flight route planning

A 2021 bachelor thesis connected to the Cognitive Science bachelor's program, focused on route planning for aircraft in dangerous environments and on how classic search compares to reinforcement learning in a realistic custom simulation stack.

The project was built as a fairly complete technical stack: multithreaded C++ for the core simulation and planners, a custom aerodynamic and physics-based flight model, random world generation, and an autopilot used both to evaluate feasibility and to constrain route output.

On the learning side, reinforcement learning models were trained with PyTorch and run through LibTorch, while OpenGL and GLFW were used for visualization and tooling around the simulation environment. Development also leaned on lower-level debugging and performance tooling such as Valgrind.

Inside that stack, A*, a local RL planner, and a global RL planner were implemented and compared for speed, behavior, path quality, and whether the proposed routes remained practically flyable. The result points toward reinforcement learning as a promising direction for robust flight route planning when strong physical constraints and dangerous environments make naive shortest-path planning insufficient.