Transformer-Enabled Leg-Based Trajectory Generation for Autonomous Aircraft Near Airports

We present a Transformer-based trajectory generation method for autonomous aircraft operating in the vicinity of airports. The proposed model is designed to generate trajectories that adhere to standard traffic pattern procedures by generating each segment, or leg, of the pattern sequentially. This leg-by-leg approach allows the model to embed leg-specific behaviors into each portion of the trajectory, thus allowing the generation of trajectories that comply with established flight conventions. While the data-driven nature of the method allows it to learn complex patterns from flight data, it lacks inherent guarantees regarding safety or dynamic feasibility. Nevertheless, the trajectories produced serve as strong candidates for subsequent verification. In particular, they may be validated using geometric techniques or optimization-based methods, enabling the integration of learning-based trajectory generation with frameworks that provide strong guarantees.