A Hong Kong University of Science and Technology (HKUST) team's new framework hands the touchdown timing call to a small reinforcement learning policy while a conventional controller keeps the drone level, in simulation only tests.
Shipboard drone recovery is one of those problems that sounds simple until you stand on a deck in a seaway. The platform pitches, heaves, and rolls on a timescale that does not match the drone's own dynamics, and the touchdown window keeps sliding. Most autonomous approaches to date either demand a precise model of the platform's motion or train a single, end-to-end neural policy that ends up brittle the moment the sea looks different from training.
A new arXiv preprint from a Hong Kong University of Science and Technology team argues there is a cleaner way to split that problem. Their framework, called WaveLander, hands the touchdown-timing decision to a small reinforcement-learning policy while leaving the job of keeping the drone level and on track to a conventional low-level flight controller. The decoupling is the actual contribution, and it is also the reason a modest RL policy can plausibly generalize to wave disturbances it has not seen in training, at least inside the simulator the authors built.
That last caveat matters. The published evidence is simulation-only: randomized wave-induced platform motions, an eight-page paper, six figures, and no flight tests at sea. Anyone reading "generalizable" or "robust" in the abstract should hold onto the fact that the framework has not yet been flight-validated. It has only been stress-tested inside a virtual wave field.
The mechanism is straightforward to describe. The platform-relative observation the RL policy receives is deliberately compact. It collapses the messy six-degrees-of-freedom motion of a heaving deck into a small set of terms the policy can act on, and the action the policy returns is just a single scalar: the vertical velocity the drone should aim for at the next instant. Everything else, from attitude stability to lateral tracking to the inner loop that fights gusts, is handled by the classical controller underneath the learned policy. WaveLander effectively reduces dynamic platform landing to a low-dimensional, timing-aware control problem, and it does so without writing down explicit switching rules between behaviors such as "hover," "descend," and "commit to touchdown."
The team behind the work is led by Ling Shi, Professor of Electronic and Computer Engineering at HKUST, Director of the Cheng Kar-Shun Robotics Institute, and Associate Director of the HKUST-DJI Joint Innovation Laboratory. Shi is an IEEE Fellow (2023), and his lab page places the work in a robotics group that has been working on control under uncertainty for years. The HKUST-DJI affiliation is contextual. DJI has not endorsed, productized, or commented on WaveLander, and nothing in the preprint implies commercial intent.
Why this matters for readers outside the reinforcement-learning bubble is the operational problem the framework is aimed at. Maritime search-and-rescue, naval logistics, and offshore inspection all want small uncrewed aircraft that can launch from a ship and recover to it without a pilot holding a radio. The hard part is not the takeoff. It is the last meter. A deck in a three-meter sea state moves with periods of a few seconds, and a drone that descends on a fixed schedule will either crash into a rising deck or get shoved sideways the moment it touches down. The two classical responses have trade-offs. Model-predictive controllers and online-optimization methods can plan around the motion but require accurate models and on-board compute for a horizon they have to keep recomputing. End-to-end RL can absorb the dynamics but tends to fail outside the disturbance distribution it was trained on, and it is hard to inspect when something goes wrong. WaveLander's pitch is a middle path: keep the well-understood controller doing what it does well, and use learning only where it has a comparative advantage, which is choosing when to commit.
There are limits the paper itself does not hide, and any reader should hold them in mind. The platform motion in the simulator is randomized, not chaotic, and "unseen disturbances" in the abstract means inside that distribution rather than across weather regimes the model has never encountered. There is no sea-trial video, no wind-over-deck measurements, no actuator-failure scenario. The framework's strongest claim is the one a simulation paper can support: that the architectural split keeps performance stable as the wave field varies within the training envelope.
The watch items for the rest of the year are concrete. First, does any version of the framework graduate from randomized simulation to hardware-in-the-loop testing on a hex-rotor or fixed-wing surrogate, the way prior wave-landing work has? Second, will the authors publish a real-sea or at-sea-condition evaluation, or will the next stop be a peer-reviewed venue where independent reviewers can probe the simulator-to-reality gap? Third, do other groups reproduce the decoupling claim on their own wave models? The submission landed on arXiv on 2026-07-01 in the robotics category. Until those questions have answers, the cleanest summary is the architectural one: split the decision you cannot model well from the decision classical control already handles well, and let a small learned policy own the part it can.