ASCENT: Transformer-Based Aircraft Trajectory Prediction in Non-Towered Terminal Airspace
Alexander Prutsch, David Schinagl, Horst Possegger · 2026
Read ASCENT: Transformer-Based Aircraft Trajectory Prediction in Non-Towered Terminal Airspace by Alexander Prutsch, David Schinagl, Horst Possegger (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-16550.txt.
Introduction
Prutsch, Schinagl, and Possegger forecast aircraft paths near small, non-towered airfields. Busy general aviation traffic raises accident risk there. Their ASCENT model is a light transformer built for the task. A motion encoder reads past flight tracks. A query-based decoder then offers several maneuver options. Domain-aware 3D coordinate handling sharpens the inputs. Tests on TrajAir and TartanAviation beat prior baselines. Ablations confirm each design choice helps.
Catalog record
| CREATOR | Alexander Prutsch, David Schinagl, Horst Possegger |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35377 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.16550
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
ASCENT: Transformer-Based Aircraft Trajectory Prediction in Non-Towered Terminal Airspace Alexander Prutsch, David Schinagl, Horst Possegger 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. I INTRODUCTION Accurate trajectory predictions are crucial for improving airspace safety systems, especially for General Aviation (GA) flights near non-towered airports. GA includes all non-commercial civilian flights, such as recreational flying, business aviation, and aerial work ( e.g., agricultural aviation). It represents the majority of the aircraft fleet; for example, around 90% of all aircraft registered in the United States fall under this category [ 11 ] . Commercial Aviation (CA), which focuses on the transport of passengers or cargo, operates under stringent safety standards: flights operate in controlled airspace according to structured procedures, with departures and arrivals conducted at towered airports under the active management of air traffic control (ATC). However, most airports are non-towered ; it is estimated that only about 4% of the roughly 20,000 airports in the US have control towers [ 17 ] . With no ATC active, pilots must rely on standard procedures, for example Federal Aviation Administration (FAA) Standard Airport Traffic Patterns [ 10 ] , and their own judgment to maintain safety. GA flights frequently operate between non-towered airports, thereby facing a higher risk of accidents . The accident rate in GA is approximately 30 times higher than in commercial aviation (CA) [ 27 ] , with fatal accidents occurring predominantly in the vicinity of airports. Fig. 1: Comparison between a CVAE-based decoder, widely applied in aircraft trajectory prediction, and our ASCENT model employing a query-based decoder. The scenario shows an aircraft take-off: our model, leveraging learnable mode queries, predicts a diverse set of trajectories that also capture the actual aircraft movement, while the baseline only predicts the common traffic pattern at the data-collection airport. The wide spread of GA, together with its relatively high accident rate, highlights the need for safety systems in non-towered airspace. One promising approach is the prediction of future aircraft trajectories, which typically relies on historical aircraft movement data combined with global context information. The predictions can help detect potential conflicts at an early stage [ 4 ] , thereby improving GA safety in non-towered terminal areas. Additionally, trajectory prediction can be applied in other tasks, e.g., to enhance aircraft traffic flow [ 21 ] and improve arrival time/delay prediction [ 42 ] . Previous research on aircraft trajectory prediction has introduced datasets from non-towered airspace [ 31 , 30 ] and developed deep learning-based methods [ 31 , 28 , 46 , 47 , 45 ] . Analogously, trajectory prediction has become a key challenge in autonomous driving, where anticipating the future positions of surrounding traffic participants strongly influences self-driving vehicle decision-making and path planning. Interest in autonomous driving has therefore driven a significant research focus on motion forecasting for self-driving vehicles [ 35 , 29 , 41 , 53 , 18 , 5 , 36 , 48 , 37 ] . While methods that originate from the autonomous driving domain target a similar task, they cannot be directly applied to aircraft trajectory prediction due to distinct challenges : unlike road traffic, airplanes do not follow strict lane networks; only runways and airspace procedures define how aircraft approach and depart airports. Moreover, aircraft operate in three-dimensional space, whereas bird’s-eye-view prediction is mostly sufficient for vehicles. Furthermore, trajectory prediction benchmarks for aviation [ 31 , 30 ] feature significantly longer scenarios and prediction horizons spanning several minutes compared to only a few seconds in the automotive domain. Although recent advances have been made, aircraft trajectory prediction remains less extensively studied than its counterpart in autonomous driving. In the self-driving domain, state-of-the-art methods typically employ transformer-based [ 38 ] modules to encode historical motion data, and combine it with structured environmental information, e.g., on lane topology [ 12 ] . Recent advances have been largely driven by improved coordinate modelings [ 53 ] , decoder designs [ 52 , 48 ] and refined processing schemes [ 52 , 36 , 37 , 48 ] . At the same time, the field of aviation trajectory prediction [ 47 , 45 ] follows different approaches, particularly in terms of decoder architectures, highlighting a divergence between the two domains. To advance trajectory prediction for General Aviation, we address the unique challenges of aircraft motion and propose a novel approach leveraging insights from state-of-the-art methods in the self-driving domain. We propose an A ircraft S equen c e En coding T ransformer (ASCENT) featuring a lightweight yet highly capable transformer-based architecture. In the decoder design, we diverge from common practices in aircraft trajectory prediction by using a mode query-based approach [ 5 ] , which better suits the problem domain compared to diffusion or variational autoencoder-based modules of related work [ 46 , 47 , 45 ] . As highlighted in Figure 1 , our approach generates more diverse outputs, thereby better covering the range of possible future actions. We evaluate our approach on the widely used TrajAir dataset [ 31 ] and provide results on the recently proposed TartanAviation dataset [ 30 ] . In summary, our main contributions include: • We propose a novel architecture for aircraft trajectory prediction, achieving a new state-of-the-art on the TrajAir dataset across multiple evaluation settings. • To the best of our knowledge, we are the first to report trajectory prediction results on the TartanAviation dataset, including cross-dataset evaluations. • We perform an extensive ablation study of our model demonstrating the effectiveness of our architecture. II RELATED WORK II-A Aircraft Trajectory Prediction Patrikar et al . [ 31 ] propose alongside their TrajAir dataset also a baseline method (TrajAirNet) for aircraft trajectory prediction in non-towered airspace. Initially, the sequence of 3D positions of each agent, given in absolute coordinates, are encoded using temporal convolutional networks [ 19 ] , before interactions between all agents are modeled using graph attention networks [ 39 ] . The airplane trajectories are decoded using conditional variational autoencoders [ 44 ] . To model the flight dynamics, the model predicts acceleration values which are then integrated to 3D positions using Verlet integration [ 40 ] . Social-PatteRNN [ 28 ] proposes a variational RNN (VRNN)-based [ 6 ] approach with a novel context modeling which is also evaluated on the TrajAir dataset. Fig. 2: Overview of our ASCENT architecture. For clarity of presentation, visualizations are reduced to 2D. First, we normalize the historical 3D flight path with respect to the current aircraft pose. Next, we encode this local history using