EMMA-LOC v0.1

Warning

Pre-implementation. This page describes proposed contracts. Class signatures, parameter types, schema fields, and behavior are subject to change before code lands. Once implementation exists, content here will be regenerated from docstrings or sourced from running tests.

Dataset. EMMA-LOC-v0.1 covers the task, target domain, input, labels, splits, and the download command; this recipe pins only the rfgen configuration. Source. rfgen (primary).

rfgen recipe

  • rfgen commit: <rfgen-commit> (frozen at release; placeholder until pinned).

  • Emitters: the rfgen emitter_zoo group, comms family (classes qpsk, ofdm), placed at varied azimuth and elevation angles. Emitter placement and arrival angles come from the rfgen scene and placement groups; EMMA does not re-specify them.

  • Receiver array: 8-element URA (uniform rectangular array), half-wavelength spacing (spacing_lambda: 0.5), via the rfgen MultiRXConfig.geometry member URA. URA is chosen so the array carries the inter-antenna phase coherence AoA (angle-of-arrival) requires and so the same scene set extends to the v1 azimuth-plus-elevation target; v0.1 scores the scalar arrival azimuth aoa_deg. Loaded into ArraySpec at load time.

  • Channel environment(s): E_URBAN (Sionna UMa, dense urban multipath) and E_RURAL (Sionna RMa, sparse rural near-LoS), pinned through the rfgen channel group. Two environments make LEAVE_ONE_ENVIRONMENT_OUT live from day one.

  • Labeler + label extraction: the rfgen label group writes per-emitter metadata; LabelExtractor maps emitters[].aoa_deg to the AoA regression target, and emitters[].los_flag to the LoS / NLoS binary label.

  • SNR (signal-to-noise ratio) regime: 0 to 30 dB, drawn from the rfgen channel.snr_db_range.

  • Sample counts: approximately 240k scenes total.

  • Splits + seeds: train [1, 160000], dev [160001, 200000], holdout [200001, 240000]. The holdout is a frozen secret seed, never published. OOD axis: LEAVE_ONE_ENVIRONMENT_OUT.

Real-capture counterpart

EMMA-REAL-OOD-v0.1. Real captures from Colosseum, POWDER, and COSMOS feed the sim-to-real gap column for localization, never the aggregate score.

Notes

  • AoA is scored as continuous regression, not binned classification; the physically meaningful quantity is an angle, and binning hides systematic bias behind class accuracy.

  • The load-bearing open risk is that rfgen’s inter-antenna phase coherence is not yet independently verified. AoA is unevaluable without correct phase across array elements. See Datasets validation.

References

  • Schmidt, “Multiple emitter location and signal parameter estimation,” IEEE Trans. Acoustics, Speech, Signal Processing 1986, DOI:10.1109/TASSP.1986.1164830. The MUSIC algorithm; grounds that AoA is a continuous angular quantity read from inter-antenna phase. (verify)

  • NVIDIA Sionna: Hoydis et al., “Sionna: An Open-Source Library for Next-Generation Physical Layer Research,” 2023, arXiv:2203.11854. The UMa and RMa ray-tracing models behind the two environments. (verify)

See Also