FlightGen: Controllable Flight Trajectory Generation with Physics-Constrained Diffusion

Controllable flight trajectory generation plays a vital role in air traffic management, especially in emergency event simulation and flight planning. Existing methods usually generate flight trajectories by numerically solving aerodynamic equations, which is computationally expensive and cannot handle complex flight patterns. Although some learning-based trajectory generation works have been proposed, they are specialized for vehicles. The airplanes fly in 3D free space and have strict waypoints regulations. The physical constraints, e.g., speed, acceleration and climb rate, are also inviolable in generation. In this paper, we propose FlightGen, which is the first learning-based controllable flight trajectory generation method. FlightGen controls the generated trajectories with pre-given airway and builds a two-stage diffusion process to adhere to physical constraints. For capturing transmission pattern of waypoints, a behavior-adaptive waypoints encoding module is designed to obtain the controlling condition. To yield realistic flight trajectories, we employ a two-stage diffusion to guide the trajectory passing through waypoints while adhering to physical constraints. Extensive experiments on two real ADS-B flight datasets show that FlightGen can not only generate trajectories highly similar to the real data but also achieves over 30x acceleration compared with existing flight simulation methods. Moreover, flight prediction models trained on the generated trajectories have performance comparable to those trained on real data.