Can drones land on moving ships without knowing where they are?

Landing a drone on a pitching ship is like trying to catch a ball thrown from a moving car. This work trains a reinforcement-learning policy that uses only the drone's onboard sensors and visual keypoints from the landing surface—no need for explicit knowledge of the platform's position or motion. Tested in realistic simulation and real-world trials across two different visual feature extractors, the method beats a conventional control baseline under "Very Rough" sea conditions and achieves autonomous landing without relying on platform-state information.