💻 computer science

DESCENT: Directed Edge Scene Encoding for Airport Surface Movement Prediction

The paper introduces DESCENT, a transformer-based architecture that leverages a Potential Reachable Set context sampling mechanism to achieve state-of-the-art accuracy in predicting airport surface movements by effectively handling heterogeneous dynamics and strict topological constraints.

Original authors: Alexander Prutsch, David Schinagl, Horst Possegger

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alexander Prutsch, David Schinagl, Horst Possegger

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The ground beneath an airplane's wheels is a place of intense, high-stakes choreography, though it rarely feels like a dance to the people watching it. While the sky is often the focus of aviation safety, the surface of an airport is a crowded, complex environment where massive machines must move with absolute precision. Here, a single aircraft might be taxiing slowly toward a terminal, while another accelerates rapidly down a runway for takeoff, and a third waits at a holding point, all within the same few minutes. Predicting where these vehicles will go next is a difficult puzzle because their movements are not random; they are strictly bound by the painted lines on the tarmac and the rules of the road. If a system can accurately guess the future path of a plane or a ground vehicle, it can warn controllers of potential collisions before they happen, turning a reactive safety system into a proactive one. This is the goal of trajectory prediction: using data to see the future of movement in a way that keeps people safe.

For years, researchers have built sophisticated computer programs to predict how cars and pedestrians move on city streets. These systems work well because traffic on roads tends to follow relatively consistent patterns; a car usually moves at a similar speed, and the environment around it doesn't change drastically from one moment to the next. However, when scientists tried to apply these same road-traffic tools to airports, they found the methods fell short. An airplane on an airport surface behaves very differently from a car. It can go from a complete stop to high speed in seconds, and the distance it needs to travel to complete a maneuver can vary wildly. A plane taxiing to a gate might only need to move a few meters, while one preparing for takeoff might need to cover several kilometers. Standard prediction tools, which look at a fixed area around a vehicle, often miss the big picture. They might focus too closely on the immediate surroundings of a slow-moving plane, ignoring the long runway it is about to use, or they might include too much irrelevant detail for a fast-moving one, confusing the computer.

To solve this, a team of researchers at the Graz University of Technology developed a new system called DESCENT. Instead of forcing the airport environment into a rigid, fixed box, their system adapts to the specific needs of the aircraft it is watching. The core of their innovation is a way of "looking ahead" that respects the physical layout of the airport. They treat the airport map not just as a collection of lines, but as a connected network of paths. When the system focuses on a specific aircraft, it asks a simple but powerful question: "Given where this plane is right now and where it is allowed to go, what parts of the airport are actually reachable?" It then gathers information only from those reachable areas. This means that if a plane is about to take off, the system automatically pulls in data from the entire runway and the approach path, even if that path is far away. If a plane is just turning a corner at a gate, the system focuses on the immediate taxiway. This approach ensures the computer has the right amount of context, no more and no less, to make an accurate guess.

The researchers tested this new method using a massive dataset of real airport movements collected from ten major airports across the United States. They compared their system against the previous best methods, which were adapted from self-driving car technology. The results showed a clear advantage for the new approach. In situations where safety is most critical—such as when an aircraft is moving near a runway or interacting with other traffic—the new system made significantly fewer errors in predicting where the plane would be fifty seconds in the future. The improvement was especially noticeable for fast-moving aircraft and complex scenarios where the path forward was long and winding. The system also proved to be efficient, running quickly enough to be useful in real-time operations, despite having to process a much richer and more detailed map of the airport than its competitors.

What makes this work particularly important is that it acknowledges the unique nature of aviation. The researchers found that simply copying what works for cars does not work for planes. By building a system that understands the specific rules and physical constraints of an airport, they created a tool that is far more reliable. The study demonstrates that when prediction models are tailored to the specific environment they operate in, they can handle the wide range of speeds and distances found in aviation much better than generic models. While the system is not a magic crystal ball, it provides a robust technical foundation for the future of airport safety. It suggests that by giving computers a smarter way to see the world around them, we can detect dangerous situations earlier and keep the flow of air travel moving smoothly and safely.

Technical Summary: DESCENT – Directed Edge Scene Encoding for Airport Surface Movement Prediction

Problem Statement
The continuous growth of commercial air traffic has increased the complexity of airport surface operations, leading to higher runway occupancy rates and a greater frequency of critical safety incidents, such as runway incursions. While motion forecasting is well-established in autonomous driving, its application to airport surface movements remains underexplored. A primary challenge in this domain is the heterogeneous dynamics of aircraft and ground vehicles. Unlike road traffic, where motion patterns are relatively homogeneous, airport operations span a vast dynamic range: from low-speed taxiing to high-speed takeoff and landing accelerations. Consequently, the spatial extent of future trajectories within a fixed time horizon varies drastically, ranging from a few meters to several kilometers. Standard map-sampling strategies used in autonomous driving (e.g., fixed-radius regions or top-k nearest elements) fail to capture this variability, often providing insufficient context for high-speed maneuvers or including excessive irrelevant data for low-speed operations.

Methodology: DESCENT Architecture
The authors propose DESCENT (Directed Edge Scene Encoding for Airport Surface Movement Prediction), a transformer-based architecture designed to handle these unique constraints. The core innovation lies in a domain-aware Potential Reachable Set (PRS) context sampling mechanism, integrated into a standard encoder-decoder pipeline:

PRS-Based Scene Sampling: Instead of fixed geometric regions, DESCENT formulates the airfield map as a graph of directed edges. It defines the PRS as the set of all airfield segments accessible from an agent's current location via valid paths in the lane graph.
- Adaptive Traversal: The sampling depth is determined by cumulative distance and segment category. For structurally coherent elements like runways, the entire segment is included to preserve long-horizon context for takeoffs and landings.
- Efficiency: This approach maintains a manageable number of scene tokens while ensuring all physically feasible paths are represented, effectively capturing context for both fine-grained maneuvers and long-range motions.
Encoding and Fusion:
- Map Encoding: Airfield segments are transformed into a segment-centric local coordinate system and encoded using a PointNet-like encoder ( $f_A$ ).
- Motion Encoding: Historical agent observations are processed via a self-attention-based motion encoder ( $f_M$ ), normalized to the most recent agent pose.
- Scene Context: Agent and map tokens are concatenated and augmented with positional and categorical embeddings. An attention-based scene encoder ( $f_S$ ) models the relational dependencies between the focal agent and the environment.
Trajectory Decoding: A detection transformer-style decoder ( $f_D$ ) utilizes learnable mode queries to generate multimodal trajectory hypotheses. The output is parameterized as a Gaussian Mixture Model (GMM), providing trajectory means, variances, and confidence scores.
Training Objective: The model is optimized using a winner-takes-all strategy combining cross-entropy for mode selection, Negative Log-Likelihood (NLL) for regression, and a novel Signed Distance Field (SDF) regularization term. The SDF loss penalizes spatial deviations from valid airfield segments, enforcing topological compliance.

Key Contributions

Potential Reachable Set (PRS) Sampling: A novel mechanism for extracting map context that adapts to the operational phase and infrastructure constraints of airport surfaces, addressing the limitations of fixed-radius sampling.
DESCENT Architecture: A trajectory prediction model integrating PRS sampling with a transformer backbone, tailored for the heterogeneous dynamics of airport operations.
State-of-the-Art Performance: Comprehensive evaluations demonstrating significant improvements over existing baselines, particularly in safety-critical scenarios.

Experimental Results
The method was evaluated on the Amelia-10 benchmark, a dataset comprising surface movement data from ten U.S. airports.

Safety-Critical Agents: In evaluations focusing on agents near potential conflict points (critical agents), DESCENT significantly outperformed the baseline Amelia-TF. For a 50-second prediction horizon, the average minimum Final Displacement Error (mFDE) improved from 105.40m (Amelia-TF) to 66.90m (DESCENT).
Random Agents: The model also demonstrated robustness on randomly selected agents, achieving lower errors across all tested airports.
Multi-Airport Generalization: A unified model trained on all ten airports achieved performance nearly matching specialized per-airport models, though some degradation was observed in highly unique operational environments (e.g., PANC, KLAX).
Ablation Studies: Comparisons against standard sampling methods (Top-K closest, fixed radius) confirmed that PRS-based sampling yields the lowest error rates, validating the necessity of topology-aware context extraction.
Efficiency: Despite the complex context sampling, DESCENT features only 3.2M parameters (compared to 89.8M in Amelia-TF). While inference latency is slightly higher due to the richer map context, the model achieves a 75% reduction in scene input size compared to using the complete map.

Significance and Claims
The paper claims that DESCENT provides a robust technical foundation for aviation-specific motion forecasting. By explicitly addressing the heterogeneous dynamics and topological constraints of airport surfaces, the approach demonstrates that deep learning methods tailored to specific domain requirements can overcome the limitations of architectures transferred directly from autonomous driving. The authors emphasize that their domain-aware sampling provides the critical long-horizon context necessary for safe navigation, particularly in safety-critical scenarios where accurate prediction of runway clearance and conflict detection is paramount. The work suggests a pathway toward semi-automated airfield operations and advanced robotic assistance for pilots, aiming to mitigate risks associated with increasing air traffic density and staffing shortages.

Technical Summary: DESCENT – Directed Edge Scene Encoding for Airport Surface Movement Prediction

More like this