Advanced automation is a key technology for enhancing the safety of ground operations amidst the increasing density of commercial air traffic. While motion forecasting is a well-studied task in autonomous driving, its application to airport surface movements remains underexplored. To enable efficient and accurate prediction in this domain, we propose DESCENT, a transformer-based architecture designed to handle heterogeneous dynamics and strict topological constraints. Our approach features a Potential Reachable Set (PRS) context sampling mechanism that adaptively collects airfield environment context across diverse operational phases. Combined with a detection transformer-based decoder, DESCENT generates accurate trajectory forecasts. Extensive evaluations on the Amelia-10 benchmark demonstrate significant performance improvements over state-of-the-art baselines. These gains are especially pronounced in safety-critical scenarios, where our domain-aware sampling provides critical long-horizon context necessary for safe navigation.
Figures & tables
Fig. 1: We illustrate two aircraft at Logan International Airport (KBOS) with future movements indicated by colored paths. Runways are shown in dark gray and taxiways in light gray. Standard map-sampling strategies – such as top- k nearest elements or fixed circular regions of interest – fail to account for the heterogeneous dynamics of airport operations. While effective in autonomous driving, a fixed region in the aviation domain may either capture insufficient context for high-speed maneuvers (dark circles) or include an excessive number of irrelevant segments (bright circles), hindering context learning and increasing computational overhead.
Fig. 2: Overview of our DESCENT architecture. We employ a potential reachable set-based airfield map sampling to obtain the map context based on the current focal agent position. Then, all agent observations O and airfield segments S are encoded using individual encoders fM and fA . Positional Pemb and type Temb embeddings are added to the tokens, to form a scene context C . The scene context is encoded using a scene encoder fS , lastly a decoder using learnable queries Q create output trajectories T and probability scores P .
Fig. 3: Illustration of our Potential Reachable Set-based context sampling (on the KSFO airport). The process begins by locating the agent’s current airfield segment (top left), followed by a graph traversal to identify reachable pathways (bottom left). We generate a sparse context (right) comprising local segments and distant but reachable long-term structures. This selection reflects the agent’s operational constraints, prioritizing high-velocity paths like runways over distant, unreachable taxiways.
Fig. 4: Comparison of random vs. critical focal agent sampling in the Amelia-10 benchmark. The top row displays the distribution of future trajectory endpoints, highlighting a more diverse distribution for critical agents. The bottom row illustrates the relationship between current agent speed and future trajectory length. Notably, critical agents show higher dynamic changes: agents with low initial speed but long future trajectories (acceleration) and high initial speed with short future paths (braking)–compared to the more linear correlation seen in random agents. Best viewed digitally with zoom.
Method
KMDW
KEWR
KBOS
KSFO*
KSEA*
KDCA
PANC
KLAX
KMSY
KJFK
Avg †
Metric
Amelia-TF [ 17 ]
66.69
91.18
81.61
-
-
105.60
170.48
137.40
87.46
102.77
105.40
mFDE@50s
DESCENT (Ours)
59.64
77.74
64.49
56.60
61.45
0 54.04
0 67.13
0 64.50
0 68.98
0 78.67
0 66.90
Amelia-TF [ 17 ]
26.41
32.88
30.79
-
-
0 39.24
0 61.35
0 50.74
0 31.41
0 39.23
0 39.01
mADE@50s
DESCENT (Ours)
24.88
31.69
26.80
23.22
24.79
0 25.46
0 28.41
0 27.43
0 27.39
0 33.56
0 28.20
Amelia-TF [ 17 ]
13.35
15.78
14.74
-
-
0 20.20
0 28.08
0 25.79
0 14.65
0 19.82
0 19.05
mFDE@20s
DESCENT (Ours)
13.60
17.20
14.25
12.40
12.68
0 15.76
0 15.44
0 15.21
0 14.88
0 19.29
0 15.70
TABLE I: Trajectory prediction results on the Amelia-10-bench test set for forecasting horizons of 50 s and 20 s. Experiments are done on the test sets, evaluating on the most critical agent per scenario. Errors reported in meters.
Method
KMDW
KEWR
KBOS
KSFO*
KSEA*
KDCA
PANC
KLAX
KMSY
KJFK
Avg
Metric
Amelia-TF [ 17 ]
27.52
52.06
51.01
40.23
65.82
47.75
86.24
89.46
29.84
55.32
54.52
mFDE@50s
DESCENT (Ours)
25.23
48.80
42.23
37.11
39.56
27.92
44.14
45.17
24.63
46.27
38.11
Amelia-TF [ 17 ]
11.80
20.96
20.47
17.05
29.94
18.93
34.90
35.84
11.42
22.43
22.37
mADE@50s
DESCENT (Ours)
11.37
21.34
18.35
15.81
17.51
13.21
20.34
20.24
10.38
20.51
16.91
TABLE II: Trajectory prediction results on the Amelia-10-bench test set for a forecasting horizon of 50 s. Experiments are done on the test sets, evaluating a random agent per scenario. Errors reported in meters.
Method
KMDW
KEWR
KBOS
KSFO
KSEA
KDCA
PANC
KLAX
KMSY
KJFK
Avg
Metric
DESCENT (Ours)
29.73
50.37
47.94
44.23
46.50
29.99
53.46
51.08
23.31
41.48
41.81
mFDE@50s
DESCENT (Ours)
13.01
21.57
20.39
18.83
20.35
14.01
24.08
22.75
10.10
17.85
18.28
mADE@50s
TABLE III: Multi-airport trajectory prediction results on the Amelia-10-bench test set for a forecasting horizon of 50 s. Experiments are done on the test sets, evaluating a random agent per scenario. Errors reported in meters.
Fig. 5: We visualize 50s DESCENT predictions (colored lines) alongside the 10s historical focal agent movement (black), surrounding agent histories (light gray), the ground truth future trajectory (thick, dark gray), and airfield map context. For clarity, predictions with confidence scores below 0.15 are omitted. The scenarios are sampled from the Amelia-10 test set showing scenes from KBOS. Best viewed digitally with zoom.
Scene Context Sampling
KBOS
KSFO
PANC
Avg
Top-K closest (50 segments)
129.82
122.75
169.92
140.83
Radius-based (0.3 km)
0 86.95
0 80.44
107.69
0 91.70
PRS-based
0 64.49
0 56.60
0 67.13
0 62.74
TABLE IV: Ablation study on scene context sampling methods. We report the mFDE@50s on the respective test sets using the most critical agent per scenario. Errors reported in meters.
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.
Alexander Prutsch, David Schinagl, Horst Possegger
Institute of Visual Computing, Graz University of Technology, Austria
Building accurate decision-support tools for next-generation air traffic control requires robust trajectory prediction models. We present a flow-matching architecture trained exclusively on historical aircraft trajectories, with no route labels or chart supervision. Trained on 1.15 million Automatic Dependent Surveillance-Broadcast trajectory windows collected over the San Francisco Bay Area, the model generates aircraft trajectory distributions that closely match historical traffic, reproducing known airspace structure around San Francisco Airport such as the shape of SFO's published NIITE FOUR departure procedure. Our model is trained directly on the native, irregular ADS-B sampling interval. Trajectory prediction is cast as sequence inpainting using a block-causal Transformer that denoises future state tokens conditioned on the observed history using Conditional Flow Matching or Denoising Diffusion Probabilistic Models. We compare our architecture against constant-velocity, deterministic-Long Short Term Memory, and Conditional Variational Autoencoders baselines. At matched parameter count, CFM outperforms DDPM by 11-26% in minADE@20, and both generative objectives surpass the CVAE baseline by 31-41%. We further show that the error degrades gracefully with prediction horizon, and the architecture remains effective when retrained on temporally decimated feeds. Lastly, we sample K independent completions, yielding spatial probabilistic occupancy estimates that can serve as input to downstream conflict-risk estimation.
Mathurin Petit, Emir Torun, Louis Brusset +2
École Polytechnique, Palaiseau, 91128, France · Technische Universität Berlin, Berlin, 10623, Germany · Mines Paris–PSL University, Paris, 75006, France +2
Accurate air traffic prediction in the terminal airspace (TA) is pivotal for proactive air traffic management (ATM). However, existing data-driven approaches predominantly rely on time series-based forecasting paradigms, which inherently overlook critical aircraft state information, such as real-time kinematics and proximity to airspace boundaries. To address this limitation, we propose AeroSense, a direct state-to-flow modeling framework for air traffic prediction. Unlike classical time series-based methods that first aggregate aircraft trajectories into macroscopic flow sequences before modeling, AeroSense explicitly represents the real-time airspace situation as a dynamic set of aircraft states, enabling the direct processing of a variable number of aircraft instead of time series as inputs. Specifically, we introduce a situation-aware state representation that enables AeroSense to sense the instantaneous terminal airspace situation directly from microscopic aircraft states. Furthermore, we design a model architecture that incorporates masked self-attention to capture inter-aircraft interactions, together with two decoupled prediction heads to model heterogeneous flow dynamics across two key functional areas of the TA. Extensive experiments on a large-scale real-world airport dataset demonstrate that AeroSense consistently achieves state-of-the-art performance, validating that direct modeling of microscopic aircraft states yields substantially higher predictive fidelity than time series-based baselines. Moreover, the proposed framework exhibits superior robustness during peak traffic periods, achieves Pareto-optimal performance under dayparting multi-object evaluation, and provides meaningful interpretability through attention-based visualizations.