ATC instruction-conditioned trajectory prediction (v2)
Checkpoints for atc-trajectory-prediction: prediction of an aircraft's next 60 s from its ADS-B history and a controller instruction known only through an imperfect speech transcript.
p(Y | O, X) ≈ Σ_Z q_φ(Z | O, X) · p_θ(Y | X, Z)
- Semantic belief model (SBM)
q_φ(Z | O, X): a distribution over interpretations (status, lateral, vertical and speed instruction) from the transcript, each word's recognizer confidence, the aircraft's own ADS-B callsign and its 30 s history. - Conditional trajectory predictor (CTP)
p_θ(Y | X, Z): a rectified flow on 18 PCA trajectory-shape coefficients. - Composition: 12 interpretations drawn from the SBM, 6 future trajectories under each.
Files
| file | contents |
|---|---|
runs-v2.tar.gz |
the main model and its baselines, first training run: final SBM, CTP, history-only predictor, transcript-conditioned predictor, mixture-density (anchors) head |
SHA256SUMS |
checksums, verified by scripts/artifacts.py fetch |
Install into a clone of the repository:
uv run python scripts/artifacts.py fetch --hub alexzongo --models-only
Then, for example:
from atc_forecast.inference import Forecaster, load_states
fc = Forecaster.load("runs/v2/sbm_jacktol_ho_s0.pt", "runs/v2/ctp_basis_atoms_zfix_s0.pt")
Results (synthetic test set, three training runs)
Energy score in metres over 60 s, lower is better, on instructions for the aircraft's own maneuver: history only 3,202; most probable interpretation 1,727 ± 2; composed 1,682 ± 5; reference interpretation given 1,556 ± 2. On a private set of 338 recorded controller transmissions matched to ADS-B, composition improves on history only by 9 %, and the addressed aircraft is ranked first among about 12 for 98 % of transmissions. Details: docs/v2/results.md in the repository.
Intended use and limits
Research on trajectory prediction under communication uncertainty. Not for operational use.
- Trained on synthetic transmissions built from real ADS-B over the San Francisco Bay Area (five days), with transcripts simulated from a fitted recognizer channel.
- The interpretation space covers heading, altitude and speed instructions. Route, conditional and procedure clearances are outside it.
- The flow with 12 Euler steps is moderately overconfident (90 % intervals cover 0.86). On recorded speech the prediction is more overconfident, and its vertical score does not beat the history-only model.
- Speed targets are compared with ground speed, and headings with track.
Citation
@software{zongo2026atc,
author = {Zongo, Alex and Petit, Mathurin and Brusset, Louis and Kam, Jordan and
Torun, Emir and Ghazanfari, Mahyar and Wei, Peng and Bayen, Alexandre},
title = {Aircraft Trajectory Prediction under {ATC} Instruction Uncertainty},
year = {2026},
note = {Software; paper in preparation}
}