WaveLander: RL framework lands drones on wave-tossed ships

Autonomous landing of unmanned aerial vehicles (UAVs) on wave-disturbed marine platforms remains a critical challenge due to stochastic platform motion, time-varying attitude, and uncertain touchdown conditions. Existing model-based methods require accurate motion prediction and online optimization, while end-to-end learning approaches suffer from high training complexity and limited interpretability. The WaveLander framework, detailed in a new preprint from researchers including Chun-Kit Li, addresses these issues by decoupling the landing decision-making from low-level flight stabilization using reinforcement learning. The RL policy takes a compact platform-relative observation—such as the vessel's heave, pitch, and roll—and outputs a single scalar vertical velocity reference. This reduces the dynamic landing problem to a low-dimensional, timing-aware control task, allowing smooth landings without explicit switching rules.

Simulation results under randomized wave-induced platform motions show that WaveLander achieves robust landing performance and generalizes to unseen disturbance conditions, outperforming both model-based and end-to-end baselines. By leveraging a hierarchical structure, the framework combines the adaptability of RL for high-level timing decisions with the reliability of conventional control for stability and tracking. This approach not only simplifies training but also enhances interpretability, as the RL policy's role is clearly defined. The work, presented in an 8-page paper with 6 figures, demonstrates the potential of hierarchical learning-based control for marine UAV recovery, a key enabler for naval operations, offshore inspections, and autonomous cargo delivery in rough seas.