ASCENT: Transformer-Based Aircraft Trajectory Prediction in Non-Towered Terminal Airspace

Alexander Prutsch; David Schinagl; Horst Possegger

cs.ROMar 17, 2026

ASCENT: Transformer-Based Aircraft Trajectory Prediction in Non-Towered Terminal Airspace

Authors: Alexander Prutsch, David Schinagl, Horst Possegger

Organizations: Institute of Visual Computing, Graz University of Technology, Austria

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.

Figures & tables

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.

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 attention blocks and combine it with global context through positional embeddings. To generate multi-modal future trajectories, we employ learnable mode queries that capture potential actions, such as landing or turning.

Panel of Figure 3 — Fig. 3: Comparison of different coordinate system normalization approaches on trajectories from the TrajAir dataset [ 8 ] . We show the current aircraft position (circular markers), along with historical (11 s) and future (120 s) trajectories. We compare global coordinates (left), object-centered normalization (middle), and our combined positional–angular normalization (right). For each method, the bird’s-eye view is shown on top, with vertical trajectories below. The global coordinate visualization shows aircraft approaching and departing from the runway, which is aligned with the x-axis. Normalizing with respect to the aircraft position moves all centers to the origin, while angular normalization aligns orientations, producing structured clusters of similar maneuvers, such as left and right turns. The top-right plot highlights a higher frequency of left turns, reflecting the left-traffic pattern at the KBTP airport.

Input: 11 s	7Days-1		7Days-2		7Days-3		7Days-4		7Days-Avg
Method	minADE 5	minFDE 5	minADE 5	minFDE 5	minADE 5	minFDE 5	minADE 5	minFDE 5	minADE 5	minFDE 5
Constant Velocity	1.79	4.08	1.90	4.31	1.92	4.30	1.82	4.16	1.86	4.21
TransformerTF [ 45 ]	1.58	3.85	1.69	4.10	1.97	4.36	1.79	4.19	1.76	4.13
STG-CNN [ 46 ]	1.19	2.35	1.36	2.70	1.33	2.67	1.17	2.29	1.26	2.50
Nearest Neighbor	3.13	2.70	1.92	1.99	3.41	2.69	2.59	2.58	2.76	2.49
TrajAirNet [ 8 ]	0.73	1.42	0.81	1.63	0.86	1.72	0.71	1.41	0.78	1.55

TABLE I: Trajectory prediction results on the dedicated TrajAir splits, along with average results across all four splits. All models predict

K{=}5

trajectories. The upper group follows the experiment setup of [ 8 ] (using

11\,\text{s}

input). The lower group compares to results from [ 12 ] , using their setup (

16\,\text{s}

input). Models are sorted by average minFDE 5 . Errors are reported in km.

Panel of Figure 5 — Fig. 4: Qualitative results on scenarios from the TrajAir [ 8 ] 111Days validation set. Visualizations show 11 s of historical aircraft movement (black), 120 s of ground truth future (gray) and predictions from our model (colored). For clearer visualization we exclude predictions with scores below 0.1.

Input: 40 s	111Days
Method	minADE 20	minFDE 20
Constant Velocity	1.85	4.16
TransformerTF [ 45 ]	1.67	3.94
Nearest Neighbor	1.97	2.70
STG-CNN [ 46 ]	1.37	2.35
S-PEC [ 51 ]	1.04	2.18

TABLE II: Prediction results on the full TrajAir test set ( 111Days ). Upper group: following the setup of [ 13 ] , all models use

h{=}40\,\text{s}

input and predict

K{=}20

trajectories. Lower group: additional results for our model using the standard setup from the dataset paper [ 8 ] . Errors reported in km.

Input: 11 s	KBTP S1		KBTP S2
Method	minADE 5	minFDE 5	minADE 5	minFDE 5
Constant Velocity	1.83	3.91	1.77	3.80
ASCENT (Ours)	0.36	0.60	0.32	0.56
Input: 11 s	KAGC S1		KAGC S2
Method	minADE 5	minFDE 5	minADE 5	minFDE 5
Constant Velocity	1.77	3.78	1.86	3.96

TABLE III: Trajectory prediction results on the TartanAviation dataset [ 9 ] . The dataset contains additional data from the KBTP airport (same location as TrajAir dataset) as well as data from the towered KAGC airport. Errors reported in km .

Decoder	Coordinate	Normalization		Positional	Flight Parameter	7Days-Avg
Architecture	System	Positional	Angular	Embeddings	Prediction	minADE 5	minFDE 5
CVAE Baseline	Local	✓	✓	✗	✗	0.68	1.33
Mode Queries	Local	✓	✓	✗	✗	0.52	0.98
CVAE Baseline	Global	-	-	-	-	0.48	0.82
CVAE Baseline	Local	✓	✓	✓	✗	0.44	0.80
Mode Queries	Global	-	-	-	-	0.38	0.60

TABLE IV: Ablation study on using different decoder modules (a CVAE-based baseline and our mode query-based model) and output prediction strategies (directly predicting xyz-positions vs. predicting flight parameters). Furthermore, we compare different coordinate modelings: global coordinates vs. local coordinates with different types of normalization. Errors given in km.

Train Dataset	Test Dataset	minADE 5	minFDE 5
TrajAir 111days	TA KBTP S1	0.69	1.25
TrajAir 111days	TA KBTP S2	0.64	1.19
TrajAir 111days	TA KBTP S1*	0.33	0.58
TrajAir 111days	TA KBTP S2*	0.36	0.60
TA KBTP S1	TrajAir 111days	0.32	0.57
TA KBTP S2	TrajAir 111days	0.33	0.57

TABLE V: Cross-dataset evaluation using the TrajAir and TartanAviation (TA) datasets. *Trajectories exceeding the TrajAir dataset recording radius are filtered out.

	Latency in ms
Method	$B=1$	$B=4$	$B=8$	$B=16$	$B=32$
TrajAirNet [ 8 ]	57	118	191	350	673
ASCENT (Ours)	11	0 13	0 14	0 18	0 17

TABLE VI: Inference latency analysis on a single NVIDIA V100 GPU across different numbers of aircraft (

B

Explore similar work

AeroCast: Probabilistic 3D Trajectory Prediction for Non-Cooperative Aerial Obstacles via Transformer-MDN Architecture
Jun 23, 2026Syed Izzat Ullah, Jose BacaAerial Robotics Multimodal Trajectory Prediction
Department of Engineering, Texas A&M University–Corpus Christi, Corpus Christi, TX 78412 USA
DESCENT: Directed Edge Scene Encoding for Airport Surface Movement Prediction
Aug 26, 2026Alexander Prutsch, David Schinagl, Horst PosseggerTrajectory Prediction Vehicle Trajectory Prediction
Institute of Visual Computing, Graz University of Technology, Austria
FlowATC: Aircraft Trajectory Prediction via Flow Matching
Sep 15, 2026Mathurin Petit, Emir Torun, Louis Brusset +2Flow Matching Air Traffic Management
École Polytechnique, Palaiseau, 91128, France · Technische Universität Berlin, Berlin, 10623, Germany · Mines Paris–PSL University, Paris, 75006, France +2

Jun 23, 2026J/K move · Enter open · S save

AeroCast: Probabilistic 3D Trajectory Prediction for Non-Cooperative Aerial Obstacles via Transformer-MDN Architecture

Syed Izzat Ullah, Jose Baca

Department of Engineering, Texas A&M University–Corpus Christi, Corpus Christi, TX 78412 USA

Aerial Robotics Multimodal Trajectory Prediction Trajectory Prediction

Open paper AskPDF arXiv

Autonomous aerial vehicles operating in shared airspace must predict the future positions of non-cooperative obstacles to plan evasive maneuvers before a collision becomes unavoidable. Unlike cooperative systems that share intent, non-cooperative obstacles such as birds, uncontrolled drones, or debris exhibit multi-modal motion that deterministic predictors cannot adequately represent. Existing methods either rely on recurrent encoders that propagate temporal information sequentially, limiting their ability to capture long-range kinematic precursors of maneuver initiation, or produce point forecasts that provide no distributional information to downstream planners. This paper presents AeroCast, a probabilistic trajectory prediction framework that combines a Transformer encoder with a Mixture Density Network output head to predict per-timestep Gaussian mixture distributions over future three-dimensional displacements. A translation-invariant consecutive displacement encoding and a calibration-oriented training objective address the input design and mode-degeneracy challenges specific to mixture-based aerial trajectory prediction. On a hybrid real-and-synthetic quadrotor corpus spanning nine motion categories, AeroCast reduces Average Displacement Error and Final Displacement Error by approximately 50% relative to the baselines over a five-second horizon, and achieves the lowest negative log-likelihood and Continuous Ranked Probability Score among all compared methods. Ablation analysis identifies velocity input and model capacity as the primary contributors to prediction quality, and positional encoding as essential for long-horizon trajectory coherence. AeroCast inference completes in 0.1ms per sample, compatible with real-time onboard deployment at 100Hz.

Abstract

Explore similar work

AeroCast: Probabilistic 3D Trajectory Prediction for Non-Cooperative Aerial Obstacles via Transformer-MDN Architecture

DESCENT: Directed Edge Scene Encoding for Airport Surface Movement Prediction

FlowATC: Aircraft Trajectory Prediction via Flow Matching