Safe Autonomous Landing of a UAV in Offset Approaches using Neural Networks

Autonomous landing remains one of the most critical and high-risk phases of Unmanned Aerial Vehicle (UAV) flight, particularly when large initial trajectory deviations must be corrected in real time. This paper proposes a robust autonomous landing framework for low-speed, fixed-wing, multi-ton class UAVs using a pilot-imitation strategy. Shallow, fully connected neural networks are employed in place of deep black-box models, enabling a lightweight architecture tailored for safety-critical flight operations. A specialized feature set, engineered from core flight mechanics principles, captures the control cues used by experienced external pilots, allowing the networks to map complex landing dynamics to precise Roll, Pitch, Yaw, and Throttle commands. The networks were trained and validated on data generated using a high-fidelity Real Time Simulator, with domain expertise guiding both feature engineering and the operational boundaries of the controllers. The experimental evaluation demonstrates that the proposed framework reliably mitigates large offset approaches, consistently stabilizing the aircraft and ensuring safe landing in diverse initial conditions. These results establish a strong baseline for computationally efficient, domain-informed autonomy, showing that shallow neural architectures, when grounded in aeronautical expertise, can deliver reliable and deployable solutions for next-generation autonomous UAV landing systems.