Goodflight: Trajectory Prediction with Diffusion Models for Non-Towered Terminal Airspace

R Raja Kumar1, Chittuluri Kavya Reddy2, D Shadik3, Doddipalli Dhanush4, Pandi Vamsi5
1Assistant Professor, Dept of Information Technology, SV College of Engineering, Tirupathi, India.
2 B.Tech , Dept of Information Technology, SV College of Engineering, Tirupathi, India.
3B.Tech , Dept of Information Technology, SV College of Engineering, Tirupathi, India.
4B.Tech , Dept of Information Technology, SV College of Engineering, Tirupathi, India.
5B.Tech , Dept of Information Technology, SV College of Engineering, Tirupathi, India.
Email:1rajakumar.r@svce.edu.in,2chittulurikavyareddy@gmail.com,4dhanush852004@gmail.com,
3shadikdudekula03@gmail.com,5vamsip158@gmail.com
Corresponding Author*: R Raja Kumar.

Abstract-Flighttrajectory prediction involves forecasting an aircraft's future path by modeling spatial and temporal patterns based on historical flight data. Techniques range from traditional state estimation and physical models to advanced deep learning approaches that capture complex spatiotemporal dependencies for accurate short-term and long-term predictions. Flight trajectory prediction in non-towered terminal airspace remains challenging due to imbalanced and diverse flight patterns. Existing systems largely focus on single-stage or short-term intention-guided prediction, limiting accuracy, diversity, and interpretability. This paper reviews current deterministic and simple generativemodels and identifies key limitations, including narrow  trajectory diversity, overfitting to common patterns such as hovering, and poor handling of rare or abnormal maneuvers that are critical for flight safety. To address these issues, a goal-oriented diffusion model framework is proposed, structured as a two-stage system: goal estimation and trajectory generation. The goal estimation stage introduces a “One-then-all” method that models a broad empirical distribution of possible flight goals through interaction-aware joint goal estimation, while the trajectory generation stage uses a transformer-based diffusion model guided by adjustable intention patterns to capture fine-grained movement uncertainties. A novel evaluation metric is also introduced to balance trajectory accuracy and diversity within social acceptability constraints Keywords: Flight trajectory prediction, non-towered terminal airspace, transformer-based diffusion model, empirical distribution, spatiotemporal dependencies, advanced deep learning approaches