Title: AirDreamer: Generalist Drone Navigation with World Models

Abstract:

Successfully maneuvering a drone through unknown and obstructed spaces demands robust generalization to novel scene configurations and a deep comprehension of how environmental structures align with the vehicle's operational limits. Existing approaches, which typically presuppose static environmental setups, depend extensively on manually engineered perception systems and rigid rule sets to direct the robot toward its destination. Such strategies are inherently tied to specific settings and struggle to generalize across different scenarios. Drawing inspiration from biological navigation mechanisms, we introduce a framework that leverages a world model to grasp environmental context, underpinned by a reinforcement-learning policy for control. To mitigate the risk of getting trapped in local minima and to promote yaw adjustments, we employ a sparse reward function that eliminates the need for hand-crafted shaping terms. Both simulation tests and real-world drone experiments demonstrate that our approach yields emergent abilities to traverse intricate, unfamiliar environments and escape local optima where conventional methods falter. On difficult maps, the system outperforms the strongest baseline by a 5.3% margin in navigation success. Moreover, the framework enables effective transfer from simulation to reality without requiring any tuning during deployment. The source code will be made publicly available.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC