Can drones learn to navigate anywhere without being told the rules?

Drones typically rely on human-engineered perception pipelines and predefined navigation rules that don't transfer across environments. This work trains a reinforcement-learning policy on top of a learned world model—effectively teaching drones to build their own mental map of the scene and decide where to go. The approach uses only sparse rewards, avoiding local optima that trap traditional methods. On real drones navigating complex mazes without retraining, it succeeds 5.3% more often than existing baselines and transfers smoothly from simulation to the real world. Code is released.