Skip to content

experiments.md

Latest commit

 

History

History
57 lines (42 loc) · 1.98 KB

experiments.md

File metadata and controls

57 lines (42 loc) · 1.98 KB
Raw

Experiment Log

This file records the current, README-facing experiment results. Long-running work-in-progress notes stay under internal_docs/.

2026-06-07 - PLATEAU NLOS Replay Suite

Goal: verify that one exported PLATEAU LOS/NLOS ray mask can be consumed by multiple downstream estimators, rather than only improving one bespoke demo.

Command:

PYTHONPATH=python:. python3 experiments/run_plateau_nlos_demo_suite.py

Inputs:

  • Mesh: data/sample_plateau.gml
  • Trajectory: 70 synthetic receiver epochs through the sample PLATEAU mesh
  • Satellites: 14 fixed azimuth/elevation directions
  • Mask contract: tow,epoch_idx,prn,is_los
  • Current generated mask: 980 rows, 641 NLOS rows, 65.4% NLOS

Outputs:

  • experiments/results/plateau_nlos_demo_mask.csv
  • experiments/results/plateau_nlos_demo_{spp,pf,fgo}_replay_summary.json
  • experiments/results/plateau_nlos_demo_suite_summary.{json,md,csv}
  • docs/assets/data/plateau_nlos_demo_suite_summary.{csv,md}
  • docs/assets/media/demos/plateau_nlos_visualization.html

Result:

Estimator Baseline RMS Mask-soft RMS RMS gain Wins
SPP 11.85 m 4.07 m 65.6% 48/68
PF 11.18 m 1.40 m 87.4% 70/70
FGO 8.10 m 0.38 m 95.4% 70/70

Interpretation:

  • The exported mask is estimator-agnostic enough to drive SPP, particle-filter, and local-FGO consumers.
  • The suite intentionally separates geometry/ray classification from solver behavior: replay scripts consume the CSV path, not the PLATEAU mesh.
  • Robust residual weighting alone is not the main result here; the geometry mask provides the decisive downstream information.

Limitations:

  • The replay uses a deterministic synthetic measurement model over a small shipped CityGML sample, not a real drive log.
  • The local-FGO replay fixes receiver clock bias to isolate position and NLOS mask effects.
  • This result supports the mask contract and demo pipeline; it is not a replacement for UrbanNav/PPC real-data evaluation.