EMMA-LOC-v0.1¶

v0.1 · proposed-contract

Task. E-LOC-AOA and E-LOC-LOS - see Tasks. What you predict. an angle of arrival (the scalar arrival azimuth at v0.1; elevation is added at v1) and a line-of-sight (LoS) or non-line-of-sight (NLoS) flag. Target domain. POSITIONING - positioning, defense, and radio-frequency (RF) drone handling.

Input¶

Raw complex multi-antenna I/Q (in-phase and quadrature), shape (num_rx, 2, N) float32, with per-scene metadata. 8-element uniform rectangular array (URA) at half-wavelength spacing.

Labels¶

Per-emitter angle of arrival in degrees (azimuth at v0.1; elevation added at v1), plus a LoS or NLoS flag.

Splits¶

train and dev are public and regenerable; holdout is private and scored by prediction submission. About 240,000 scenes.

Get it¶

emma download --dataset EMMA-LOC-v0.1 --split dev --out ./scenes

Metric¶

MeanAngularError for E-LOC-AOA and BalancedAccuracy for E-LOC-LOS. See Metrics.

What this tests¶

Angle of arrival is scored as continuous regression, not binned classification, because the physically meaningful quantity is an angle; binning would hide a systematic angular bias behind class accuracy.

How it is made¶

Generated by rfgen. The exact configuration (commit, emitters, channel, label extraction, seeds) is in the developer recipe.