A model called DESCENT reported lower trajectory-displacement errors than Amelia-TF when the Amelia-10 benchmark focused on critical agents over a 50-second forecast. DESCENT's average mFDE was 66.90 metres, compared with 105.40 metres for Amelia-TF, while its average mADE was 28.20 metres, compared with 39.01 metres. The study's two displacement-error measures, mFDE and mADE, are reported in metres, so lower numbers indicate closer predictions in this benchmark comparison.
The result was not limited to the critical-agent setup. For randomly selected agents at 50 seconds, DESCENT's average mFDE was 38.11 metres against 54.52 metres for Amelia-TF, and its average mADE was 16.91 metres against 22.37 metres. The study reports both random focal-agent and criticality-based focal-agent evaluations, making the two comparisons different views of the same forecasting task.
At a shorter horizon, the critical-agent comparison also favored DESCENT on mFDE: 15.70 metres at 20 seconds versus 19.05 metres for Amelia-TF.
A model built around airfield context
Those results come from Amelia-10, a benchmark containing recordings from ten U.S. airports. The study evaluates DESCENT against baselines with particular attention to safety-critical scenarios and reports both random focal-agent and criticality-based focal-agent setups.
DESCENT's central design choice is a domain-aware Potential Reachable Set, or PRS, scene-sampling mechanism. It is paired with a detection-transformer-inspired decoder that generates multimodal trajectory predictions, or multiple possible future paths for an agent. Together, those components define how the model selects airfield context and turns it into candidate trajectories.
Under the benchmark protocol, data are sampled at 1 Hz, the model receives 10 seconds of past observations, predicts 50 seconds ahead and produces four trajectory modes. Training and validation use half of the dataset, while reported results are calculated on the complete test set.
The training objective combines winner-takes-all mode selection, cross-entropy classification, negative-log-likelihood trajectory regression and SDF-based map-compliance regularization. In plain terms, that objective asks the system to choose among modes, classify them, fit trajectories and account for map compliance.
The gain came with a latency trade-off
In a joint multi-airport evaluation at 50 seconds, DESCENT's average mFDE was 41.81 metres and its average mADE was 18.28 metres.
In the reported ablation, PRS-based sampling had the lowest average mFDE of the three tested approaches: 62.74 metres, versus 91.70 metres for radius-based sampling and 140.83 metres for top-k sampling. That ranking is a descriptive benchmark result, not a causal finding about airport operations.
DESCENT was much smaller by parameter count, with 3.2 million parameters compared with 89.8 million for Amelia-TF, and the study reports a 75% reduction in scene-input size. But the compact model was slower in the reported inference comparison. On an NVIDIA L40 GPU, DESCENT took 32, 68 and 132 milliseconds for scenes with 32, 64 and 128 agents, respectively, compared with 16, 36 and 76 milliseconds for Amelia-TF.
Taken together, the benchmark numbers show a trade-off: lower reported displacement errors and fewer parameters, but higher latency than the comparison model. Because the study evaluates model performance on Amelia-10, these figures should be read as benchmark evidence rather than proof of improved real-world airport safety.
The document is an arXiv version 2 preprint dated 27 August 2026. The work was partially funded by the Austrian Research Promotion Agency under the Take Off project SAFER, project 894164.
Paper data and sources
Original title: DESCENT: Directed Edge Scene Encoding for Airport Surface Movement Prediction
Authors: Alexander Prutsch, David Schinagl, Horst Possegger
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text