Autonomous UAV Landing on Stochastic Maritime Targets

Abstract

Reliable autonomous recovery of Unmanned Aerial Vehicles (UAVs) on moving maritime platforms remains a critical challenge, primarily due to complex, stochastic deck motion, particularly vertical heave, and unpredictable environmental disturbances. Existing Reinforcement Learning (RL) approaches often simplify this environment, limiting their real-world applicability. This thesis investigates the robustness trade-offs of RL-based guidance controllers under realistic, high-dynamicity maritime conditions. We benchmarked a classical Proportional-IntegralDerivative (PID) controller against two RL architectures trained using Soft Actor-Critic (SAC) in a high-fidelity PyBullet simulation: a Full RL 3D controller and a novel Hybrid RL 1D controller, which strategically applies RL only to the critical, stochastic vertical (heave) axis. The results demonstrate that the Hybrid RL 1D architecture (86.6% success rate) achieved superior overall robustness and efficiency. Notably, the RL controllers dramatically reduced average landing time (RL_1D: 3.31 s vs. Baseline: 11.51 s), though the classical PID baseline maintained higher horizontal precision (Err𝑋𝑌 of 0.17 ± 0.17 m ). The Hybrid RL 1D maintained a superior success rate up to 89% in high sea states (SS7) and exhibited greater resilience to sensor noise. However, a critical limitation was identified: both RL-based policies experienced a pronounced performance collapse under strong, untrained wind disturbances, a regime where the non-adaptive classical PID baseline proved unexpectedly stable. These findings confirm the benefits of hybrid control for maximizing robustness and highlight that the system’s ability to handle wind disturbance rejection remains a significant, unresolved shortcoming for current RL guidance systems.