PaperlayerView exact revision on arXiv

Source-linked AI summary

DESCENT: Directed Edge Scene Encoding for Airport Surface Movement Prediction

Alexander Prutsch, David Schinagl, Horst Possegger

arXiv:2608.26002v1cs.RO

TL;DR

Airport surface movement forecasting remains less explored than autonomous-driving forecasting, despite denser operations, heterogeneous dynamics, and strict topological constraints that affect safety. DESCENT combines PRS-based adaptive airfield-context sampling with a transformer architecture for multimodal trajectory prediction, and evaluations on Amelia-10 report significant improvements, especially for safety-critical scenarios.

  • Problem

    Airport surface movement forecasting is underexplored relative to autonomous driving, although increasing traffic and heterogeneous, topology-constrained movements create important safety concerns.

  • Method

    DESCENT uses Potential Reachable Set sampling to adaptively select feasible airfield context and supplies it to a transformer-based multimodal trajectory prediction architecture.

  • Results

    Evaluations on Amelia-10 demonstrate significant performance improvements over previous work, particularly for safety-critical agents and long-horizon scenarios.

  • Takeaways & Limitations

    Domain-aware sampling provides a technical foundation for aviation-specific motion forecasting across varying spatial scales of airport movements.

Abstract

from arXiv · show

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.

I. INTRODUCTION

Airport surface traffic is becoming denser and more safety-critical, while airport movements exhibit dynamics and spatial scales that differ substantially from road traffic. DESCENT addresses these challenges with PRS-based context sampling integrated into a trajectory prediction model, achieving especially strong results for safety-critical agents.

  • Rising airport surface movements increase runway occupancy, traffic-management complexity, and the frequency of safety-critical situations such as runway incursions.
  • Airport movements range from slow taxiing to high-speed takeoff and landing, with trajectories spanning meters to kilometers within the same prediction horizon.Rapid accelerations, including rolling starts, make motion dynamics highly context-dependent.
  • DESCENT represents airport topology as directed edges and uses Potential Reachable Set sampling to extract relevant airfield context while controlling scene-token count.Traversal follows connected lane segments by semantic type, retaining complete structurally coherent elements such as runways.
  • DESCENT integrates the sampling strategy into a trajectory prediction model and achieves state-of-the-art performance on Amelia-10.
  • Comprehensive Amelia-10 evaluations show benefits particularly for safety-critical agents, while maintaining low inference latency and capturing long-horizon scene context.

II. RELATED WORK

Motion forecasting research has been driven largely by autonomous-driving methods, but airport surface operations require specialized handling because aircraft can traverse much larger spatial extents during takeoff or landing.

  • Autonomous-driving research provides sophisticated components and architectures that form a foundation for motion forecasting in other domains.
  • Standard methods commonly sample map elements from a fixed spatial region around the agent before encoding agent, map, and scene interactions.
  • Airport surface operations limit the effectiveness of standard methods because takeoff or landing may require context spanning the entire airport.

B. Aircraft Trajectory Prediction

Airport surface prediction differs from airborne trajectory prediction because surface operations combine dense interactions, discrete routing, and strong geometric constraints. DESCENT addresses this setting with directed-edge scene encoding, PRS sampling, and multimodal transformer forecasting.

  • Amelia-10 is a large-scale benchmark built from selected data collected at ten U.S. airports, with Amelia-TF as its baseline model.
  • Airborne trajectory prediction typically involves sparse interactions, continuous flight dynamics, and comparatively weak map constraints.
  • Airport surface operations instead involve dense traffic, frequent interaction, discrete routing structures, and strong geometric and topological constraints.
  • The architecture combines PRS-based map sampling with separate encoders for agent observations and airfield segments, scene encoding, and a query-based trajectory decoder.
  • DESCENT encodes historical observations and airfield map data to produce k multimodal future trajectories with associated confidence scores.Its PRS sampling supplies physically grounded context to a transformer-based backbone focused on feasible maneuvers.

A. Airfield Segment Sampling

PRS sampling adaptively selects airfield segments reachable through valid lane-graph paths, preserving relevant long-range context while keeping the model input manageable. The procedure combines semantic segment processing, graph traversal, and operational constraints.

  • The Potential Reachable Set is the set of airfield segments accessible from an agent’s current location through a valid lane-graph path.
  • PRS sampling adapts context selection to infrastructure constraints instead of using fixed-size regions or a predefined number of nearest tokens.This preserves trajectory-relevant information while keeping input token counts manageable.
  • Reachable distance is bounded by lane-type-specific thresholds rather than instantaneous vehicle metrics.This reflects cases where a stationary aircraft at a runway threshold may travel farther than a slower aircraft taxiing toward a terminal.
  • The sampler traverses a lane graph from the current segment to obtain sparse context containing local segments and distant reachable structures.Runways receive priority over distant unreachable taxiways, supporting long-horizon maneuvers.
  • Raw map data are reorganized into semantically meaningful segments by merging consecutive edges belonging to the same runway or taxiway.

B. Airfield Segment Encoding

DESCENT encodes agent histories and airfield segments into a unified scene context, then uses attention and mode queries to generate multimodal trajectory distributions.

  • B. Airfield Segment Encoding: Airfield segments are normalized into local coordinates and encoded into one token per segment, producing map context Senc ∈ R^Ns×D.A PointNet-like encoder processes Ps sampled 2D points for each of Ns segments.
  • B. Airfield Segment Encoding: Historical observations are transformed into the latest aircraft-centered frame, temporally self-attended, and pooled into compact per-agent motion tokens.This normalization focuses encoding on motion patterns independent of global position and orientation.
  • B. Airfield Segment Encoding: Agent and map tokens are concatenated into scene context and augmented with pose-based positional embeddings before scene-level attention models their dependencies.The embeddings encode 3D position and yaw for each token’s local coordinate system.
  • B. Airfield Segment Encoding: Learnable mode queries use factorized cross-attention over the encoded scene to parameterize a Gaussian Mixture Model for future trajectories.MLP heads output trajectory means, variances, and confidence scores for k intention modes.

IV. EXPERIMENTAL SETUP

The experiments train DESCENT on heterogeneous Amelia-10 airport data using specified transformer settings, random or critical sampling, and topology-aware regularization.

  • IV. EXPERIMENTAL SETUP: Amelia-10 contains recordings from ten U.S. airports with varied runway layouts, terminal configurations, and traffic densities.Official preprocessing and dataset-author configurations generate the training, validation, and test splits.
  • IV. EXPERIMENTAL SETUP: The model uses D = 128, four attention blocks in each encoder, and decoder cross-attention depth three.These settings are selected through empirical validation for the encoders and decoder.
  • IV. EXPERIMENTAL SETUP: Experiments compare random and critical focal-agent sampling in the Amelia-10 benchmark.Critical agents exhibit more diverse future endpoints and stronger acceleration or braking dynamics than random agents.
  • IV. EXPERIMENTAL SETUP: Agent inputs include 3D position, speed, and yaw, while the model distinguishes ground vehicles from aircraft and taxiways, runways, and hold lines.The setup samples Ps = 20 points per airfield segment and constrains segment lengths for balanced granularity.
  • IV. EXPERIMENTAL SETUP: Training uses batch size 128 and an initial learning rate of 10^-4, reducing the rate by one order of magnitude when validation loss plateaus.Training concludes once the learning rate reaches 10^-8.
  • IV. EXPERIMENTAL SETUP: A signed-distance-field loss penalizes deviations from predefined airfield segments to promote topologically feasible trajectories.The SDF is rasterized at 5 m resolution.

C. Metrics

DESCENT is evaluated with benchmark displacement metrics that select the best-matching trajectory among k predicted hypotheses.

  • C. Metrics: mADE measures mean Euclidean distance across future time steps, while mFDE measures endpoint distance only.Both metrics select the hypothesis closest to ground truth from the k predictions.

V. RESULTS AND DISCUSSION

DESCENT is compared with Amelia-TF across critical and random agents, airport settings, horizons, and context-sampling ablations, with strongest benefits in long-horizon critical cases.

  • A. Single Airport Evaluation: The evaluation includes single-airport, cross-airport, context-sampling ablation, and latency studies using reproducible Amelia-TF comparisons.The Amelia-TF baseline is rerun with official code and pretrained weights where available.
  • A. Single Airport Evaluation: DESCENT significantly outperforms Amelia-TF for safety-critical agents over the 50 s prediction horizon.The authors associate this result with capturing long-range contextual dependencies and complex scene interactions.
  • A. Single Airport Evaluation: Table I reports critical-agent results at 50 s and 20 s, while Table II reports random-agent results at 50 s.Both tables report errors in meters on Amelia-10 test sets.
  • A. Single Airport Evaluation: Reported Amelia-TF values may rely on the original paper because checkpoints for the official refactored codebase were unavailable.This caveat applies to values taken from reported results.
  • A. Single Airport Evaluation: At 20 s, the performance gap narrows and results are largely comparable because shorter-term predictions are less challenging.The benefit of sparse heterogeneous context is also less pronounced when agents remain near their current positions.
  • A. Single Airport Evaluation: Per-airport errors correlate with ground-truth trajectory extent, with KMDW averaging approximately 430 m and PANC and KLAX exceeding 700 m.The reported relationship indicates that error metrics vary with the operational scale of each airfield.
  • A. Single Airport Evaluation: For random focal agents, DESCENT retains significant long-term advantages in mFDE and achieves better or comparable mADE across evaluated airports.The random-agent task generally has lower errors than the critical-agent setup.

B. Multi Airport Evaluation

A single model trained jointly across Amelia-10 airports achieves high accuracy overall, nearly matching specialized per-airport models, though unique operational patterns favor specialization at the hardest airports.

  • The unified model nearly matches specialized per-airport models across airports and surpasses individual models at KJFK and KMSY.The authors attribute the gains at KJFK and KMSY likely to greater training-scenario diversity.
  • Performance degradation is most pronounced at PANC and KLAX, where specialized models can be favorable because operational patterns are highly unique.

C. Ablation Study

The ablation compares nearest-segment, radius-based, and PRS-based context sampling, with PRS achieving the best mFDE results across airports.

  • Figure 5 visualizes 50 s DESCENT predictions against 10 s histories, ground-truth futures, surrounding-agent histories, and airfield map context.Predictions with confidence scores below 0.15 are omitted.
  • PRS-based sampling achieves the best mFDE results across easy KSFO, medium KBOS, and challenging PANC.The comparison evaluates top-k closest lane segments, all segments within a predefined radius, and Potential Reachable Set-based sampling.

D. Latency Analysis

DESCENT substantially reduces model size while using richer environmental context, achieving practical inference latency and improved accuracy; PRS runtime depends on scene-segment selection.

  • DESCENT uses 3.2M parameters versus Amelia-TF’s 89.8M, enabled by a smaller latent dimension and more compact map encoder.The latent dimension is D = 128 for DESCENT versus D = 256 for Amelia-TF.
  • Inference latency is 32 ms, 68 ms, and 132 ms for predicting 32, 64, and 128 agents, respectively.Amelia-TF reports lower latencies of 16 ms, 36 ms, and 76 ms for the same agent counts, while using minimal map context.
  • PRS hyperparameters affect runtime by determining the number of airfield segments and the complexity of the associated processing functions.
  • The method reduces scene input size by 75% relative to the complete map while prioritizing comprehensive scene input for predictive performance.
  • Additional flight-state constraints could reduce input tokens and further optimize latency.Examples include whether an aircraft has landed or is scheduled for departure.
Loading 2608.26002v1…