Abstract
Accurate trajectory prediction can improve General Aviation safety in non-towered terminal airspace, where high traffic density increases accident risk. We present ASCENT, a lightweight transformer-based model for multi-modal 3D aircraft trajectory forecasting, which integrates domain-aware 3D coordinate normalization and parameterized predictions. ASCENT employs a transformer-based motion encoder and a query-based decoder, enabling the generation of diverse maneuver hypotheses with low latency. Experiments on the TrajAir and TartanAviation datasets demonstrate that our model outperforms prior baselines, as the encoder effectively captures motion dynamics and the decoder aligns with structured aircraft traffic patterns. Furthermore, ablation studies confirm the contributions of the decoder design, coordinate-frame modeling, and parameterized outputs. These results establish ASCENT as an effective approach for real-time aircraft trajectory prediction in non-towered terminal airspace.
Introduction
In General Aviation (GA), which encompasses recreational flying and small business jets, the accident rate is approximately 30 times higher than in commercial aviation. A significant contributing factor is that 96% of U.S. airports are "non-towered," lacking Air Traffic Control (ATC) to sequence arrivals and departures. Pilots must rely on standardized traffic patterns and visual observation to maintain separation.
While trajectory prediction is well-studied in autonomous driving, aircraft present unique challenges: three-dimensional freedom of movement, the absence of physical lane markers, and the need for longer prediction horizons (e.g., 2 minutes) to allow for safe maneuvering. ASCENT (Aircraft Sequence Encoding Transformer) addresses these challenges by moving from slow generative models to a fast, discriminative architecture.
Methodology
ASCENT’s architecture is designed to adapt automotive trajectory prediction "best practices" to the aviation domain. The model consists of three primary components:
- Domain-Aware Coordinate Modeling: Raw GPS coordinates are normalized relative to the aircraft’s current position and orientation (yaw and pitch). Global context, specifically the runway location, is re-introduced via a 3D positional embedding.
- Transformer-Based Motion Encoding: Self-attention blocks process historical flight data, allowing the model to prioritize critical maneuvers, such as turning onto the "base leg" of a landing pattern, over steady-state flight.
- Query-Based Decoding with Kinematics: Instead of using a Variational Autoencoder (CVAE), ASCENT utilizes learnable mode queries that represent distinct potential intents (e.g., landing vs. go-around). The model predicts flight parameters—velocity, yaw, and pitch—rather than raw XYZ coordinates. These parameters are then passed through a physics-based kinematic layer to generate 3D paths.
Experimental Results
The model was evaluated on the TrajAir and TartanAviation datasets, which contain extensive ADS-B data from non-towered and towered airports.
Accuracy and Multi-modality
ASCENT significantly outperforms previous state-of-the-art models. On the TrajAir dataset, it achieved a reduction in Final Displacement Error (FDE) of over 50% compared to baselines like TrajAirNet. Figure 1 illustrates ASCENT's ability to generate diverse, physically plausible maneuver hypotheses that closely align with ground truth, whereas traditional CVAE approaches often produce "blurry" or less accurate averages.
Efficiency and Latency
Real-time safety applications require low latency. ASCENT processed batches of 32 aircraft in 17ms, a 40x speedup over the TrajAirNet baseline (673ms). This efficiency makes it suitable for deployment on mobile devices or tablets within the cockpit.
Conclusion and Future Work
ASCENT demonstrates that combining transformer-based encoding with parameterized, query-based decoding provides a robust solution for aircraft trajectory prediction. By predicting physical parameters rather than raw coordinates, the model ensures stable and realistic outputs.
Future research will focus on incorporating explicit map knowledge (e.g., runway lanelets) and external context such as weather conditions and ATC communications. Additionally, modeling the "social" interactions between multiple aircraft in a shared pattern remains a key area for development to further enhance collision avoidance capabilities.
Reference: Prutsch, A., Schinagl, D., & Possegger, H. (2026). ASCENT: Transformer-Based Aircraft Trajectory Prediction in Non-Towered Terminal Airspace. arXiv:2603.16550v1.
Summary updated