DESCENT: Directed Edge Scene Encoding for Airport Surface Movement Prediction

Following the airport surface trajectory prediction task defined in the Amelia framework, DESCENT takes historical observations over a past horizon of $H_p$ time steps together with airfield map data, and predicts $k$ future trajectory hypotheses with associated confidence scores for a focal agent over a future horizon $H_f$. The multimodal formulation accounts for the inherent uncertainty of taxiing maneuvers, ensuring high coverage of potential future actions. The architecture is built on the following core components:

• Potential Reachable Set (PRS) Sampling: We define the PRS as the set of all airfield segments accessible from an agent's current location via a valid path in the lane graph. Unlike fixed regions of interest or top-k nearest tokens, this selects context adaptively from infrastructure constraints, keeping token counts manageable while preserving the information relevant for prediction.
• Directed Edge Map Representation: As a preprocessing step, raw map data is reorganized into semantically meaningful airfield segments. Consecutive graph edges belonging to the same structural element are merged while intersection nodes and edge directions are preserved to maintain topological consistency, and a type-dependent maximum segment length keeps granularity balanced.
• Type-Aware Traversal Depth: Reachable distance is bounded by lane-type-specific thresholds rather than instantaneous vehicle metrics, because current velocity poorly predicts future path length — an aircraft stationary at a runway threshold will cover far more ground over the next horizon than one taxiing toward a terminal. If a runway segment is involved, the entire runway is included to cover long-horizon takeoff and landing trajectories.
• Airfield Segment Encoding: Sampled segments $S \in \mathbb{R}^{N_s \times P_s \times 2}$ are normalized into a segment-centric local frame and encoded with a PointNet-like airfield encoder $f_A$, producing one token per segment.
• Motion Encoding: Agent observations $O \in \mathbb{R}^{N_a \times T_h \times D_m}$ are transformed into a local frame defined by the most recent pose, projected into feature space, processed by temporal self-attention, and pooled over time into a compact per-agent token by a lightweight motion encoder $f_M$.
• Scene Encoding: Agent and map tokens are concatenated into a scene context $C \in \mathbb{R}^{(N_a+N_s) \times D}$, each augmented with a positional embedding $P_\text{emb}$ of its local frame pose and a categorical embedding $T_\text{emb}$ for segment type and agent class. An attention-based scene encoder $f_S$ then models the relational dependencies between the focal agent and all other scene elements.
• Trajectory Decoding: A decoder $f_D$ inspired by detection transformers applies factorized cross-attention from $k$ learnable mode queries $Q$ onto the encoded scene. The updated mode tokens parameterize a Gaussian Mixture Model, and shallow MLP heads emit trajectory means, variances, and confidence scores.
• SDF Map-Compliance Regularization: Because airport surface motion is constrained to predefined segments, training adds a signed distance field term. A rasterized SDF at 5 m resolution stores the distance to the nearest airfield segment; predicted points are projected onto this grid and off-track predictions are penalized, promoting topologically feasible trajectories.

To extract the environmental context for a given vehicle, we first identify its current segment. Candidates are selected by Euclidean distance and refined by a heading consistency constraint, which avoids incorrect associations in dense intersection areas: if the closest segment's orientation deviates significantly from the vehicle's heading, the next $n$ candidates are evaluated, defaulting to the closest segment only if no angular match is found. From the current segment, a depth-first traversal of the lane graph — constrained by domain-specific length thresholds — enumerates all admissible paths. These are precomputed offline and stored as segment-ID metadata, so the map can be filtered at test time with minimal computational overhead. A backward traversal from the focal agent's current position additionally incorporates historical path context.

In safety-critical scenarios and at the long 50 s horizon, DESCENT significantly outperforms the Amelia-TF baseline, underscoring its ability to capture long-range contextual dependencies and complex scene interactions. At the shorter 20 s horizon the gap narrows and results are largely comparable, as short-term predictions are inherently less challenging and sparse heterogeneous context matters less when agents stay within a limited spatial vicinity. Per-airport accuracy correlates with the average spatial extent of ground-truth trajectories: future trajectories at KMDW average roughly 430 m, whereas airports with higher absolute errors such as PANC and KLAX exceed 700 m.

With random focal agents the task is less challenging, resulting in generally lower errors than in the critical-agent setup. DESCENT nevertheless retains a clear advantage, particularly for long-term prediction as measured by mFDE, and achieves better or comparable mADE across all evaluated airports — further validating the robustness of the PRS-based context sampling.

The jointly trained model surpasses the specialized per-airport models at KJFK and KMSY, likely due to the increased diversity of training scenarios, and nearly matches them elsewhere. Degradation is most pronounced at PANC and KLAX, two of the most challenging airports in the dataset, indicating that airfields with highly unique operational patterns can still favor specialized models.

PRS-based sampling outperforms both standard strategies at every airport, reducing the average mFDE@50s by 32% relative to radius-based sampling and by 55% relative to top-k nearest segments.

The clips below roll out the full $50\,\text{s}$ prediction horizon step by step for scenarios from the KBOS test set. Black dots mark the observed $10\,\text{s}$ history of the focal agent, the thick gray line the ground truth future, and the colored lines the trajectory hypotheses produced by DESCENT.

Multimodal forecasting cannot resolve every intent. The clips below show scenarios in which none of the $k = 4$ hypotheses tracks the executed maneuver — typically when an agent takes an unexpected branch at an intersection, or when its dynamics change abruptly within the horizon.