NeuVNav: Neuro-Evolving Visual Navigation for Indoor Lighter-Than-Air UAVs

Autonomous indoor navigation of lighter-than-air (LTA) unmanned aerial vehicles presents a distinct set of challenges absent from rotary-wing platforms: the low mass-to-volume ratio produces an inherent timing gap between visual boundary detection and effective heading change that renders reactive threshold controllers insufficient in isolation. This paper presents NeuVNav, an AI navigation system combining a two-loop Spiking Neural Network (SNN) reactive controller with an online Neuro-Evolution (NE) parameter optimizer for GPS-denied, indoor boundary patrol on the h-aero® semirigid LTA platform. The SNN encodes actuation timing as five evolvable thresholds operating on camera and ultrasonic sensor input. The NE optimizer applies an Evolution Strategy with a 9-component physics-aware fitness function to converge controller parameters without hardware interaction. Simulation results demonstrate consistent fitness convergence within 20 generations, and the implementation achieves full test coverage across the unit tests executable on standard hardware without aerial testbed access. The system is released as an open-source Python library under the ENFIELD programme (EU grant 101120657), fulfilling the Open Science requirement of the grant and providing a reproducible baseline for SNN-based LTA navigation research.