Skip to content

Draft the combined EKF/UKF/RBPF anomaly-aiding paper for ION NAVIGATION - #420

Merged
jbrodovsky merged 3 commits into
mainfrom
claude/anomaly-papers-ukf-rbpf-ekf-4d98d2
Oct 2, 2026
Merged

jbrodovsky merged 3 commits into
mainfrom
claude/anomaly-papers-ukf-rbpf-ekf-4d98d2

Conversation

@jbrodovsky

@jbrodovsky jbrodovsky commented Oct 2, 2026 •

Copy link
Copy Markdown
Owner

What

Adds papers/anom_combined/, a first complete draft of an ION NAVIGATION Regular Paper (ION's IONconf-v2 template, APA 7). It combines the ITM 2026 UKF paper and the Pacific PNT 2026 RBPF paper and adds the EKF. It is 19 pages (the limit is 20), with 8 figures and 9 tables. The PR also includes a revised data paper (see below).

This PR adds only paper material. Nothing under core/, sim/, geonav/, analysis/ or conf/ changes, and no new simulation runs were made.

Why

The two conference papers' results do not hold. The draft therefore explicitly corrects and supersedes them (Section 2), and reports the corrected, paired evaluation from the September runs:

  • Smartphone gravity and magnetic readings: no benefit to the EKF or UKF (median RMSE ratio 1.000–1.001, all p ≥ 0.13). The mechanism is quantified: the OS pins the gravity norm to a device constant, and the vehicle's field dominates the magnetometer.
  • Simulated RM3100 magnetometer: about a 10% gain for both Kalman filters.
    • EKF: ratio 0.891 [0.866, 0.907], better on 25 of 27 drives.
    • UKF: ratio 0.901, better on 26 of 27.
  • Simulated ADXL355 gravimeter: no gain.
  • RBPF: reported as a negative result under 60 s GNSS. It diverges with or without aiding. Its reported 1σ is about 74 m against a median error of about 880 km, which points to particle deprivation.

Reviewer notes

  • Conference-era defects were verified, not recalled. Each defect in Section 2 was confirmed by reading the source at the two conference-paper commits:

    • 4a4e3de (UKF paper)
    • 4ea73af (RBPF paper)

    The defects: anomaly units, radian latitude, the missing free-air term, the Eötvös sign and radii, the WMM subtracted in tesla, and the geonav event stream diverging from core's. The divergence matters only for the RBPF paper: its aided runs lacked the heading update the unaided runs had.

  • Every number is recomputed from the run CSVs. The scripts in scripts/ read STRAPDOWN_DATA (default: the main checkout's data/), and build_tables.py is deterministic. The numbers match the dissertation where the dissertation is self-consistent. Where it is not, NOTES.md lists each discrepancy; one example is the UKF α, which was 0.1 for these runs, not 1e-3.

  • Open item, provenance: data/output_real2 (the smartphone arm) has no committed recipe, and conf/real/rbpf_*.toml no longer parses. Regenerate that arm from a committed config before submission. This is in NOTES.md.

  • Data paper (papers/mems_nav_data/, commit bec1291): the rejected MEMS-Nav data paper, retargeted to Data in Brief. It adds device and trajectory statistics, GNSS-accuracy and gravity-norm characterization, and EKF/UKF/RBPF benchmark baselines. scripts/build.py regenerates every table and figure. Author decisions (Zenodo release, consent/ethics, mounting details, CRediT) are marked TODO in the text and listed in its NOTES.md. App device identifiers are deliberately not reproduced.

  • Review follow-up (commit 99552a7):

    • The map gradient is now taken from the bilinear cell containing each point (1.44 mGal/km and 6.37 nT/km).
    • The noise/gradient bound is presented as a heuristic, not a bound.
    • The RBPF diagnostics are now scripted.
    • The RBPF particle counts are corrected (2,000 with full GNSS, 1,000 degraded), and provenance is qualified: the configs the runs read are at 2a40f1e, and the binary's revision is unverified.
  • Build: latexmk -pdf main.tex builds with no undefined references or citations. Ruff check and format pass on the scripts.

🤖 Generated with Claude Code

Adds papers/anom_combined: a NAVIGATION Regular Paper draft (ION's IONconf-v2
template) that combines the ITM 2026 UKF and Pacific PNT 2026 RBPF conference
papers, adds the EKF, and explicitly withdraws their results. The defects behind
those results (anomaly units, radian latitude, missing free-air term, Eotvos
sign, WMM in tesla, mismatched event streams, diverged baselines) were confirmed
at the revisions that produced them, 4a4e3de and 4ea73af.

Every number is computed from the run outputs by scripts/ (stats, tables,
figures, phone-sensor mechanism), which read STRAPDOWN_DATA and are
deterministic. NOTES.md lists the open items, including the unrecorded
provenance of the smartphone-arm runs and where the dissertation's figures
differ from the recomputed ones. No source, config or analysis code changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Copilot AI balanced review requested due to automatic review settings October 2, 2026 19:48

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copilot review overview

🟡 Changes recommended

The gradient calculation, reproducibility claims, and scientific interpretations need correction.

Review effort: Balanced
Findings: 1 High severity · 1 Medium severity · 7 Low severity

Open (9)
What changed in this PR

Adds a journal-paper draft presenting corrected anomaly-aiding evaluations for strapdown-rs. Navigation code is unchanged, and no new simulations were run.

Changes:

  • Combines and supersedes the earlier conference results, adding the EKF comparison.
  • Adds scripts and generated tables for analyzing existing runs.
  • Documents experimental limitations, provenance gaps, and rebuild instructions.
File Description
papers/​anom_combined/​tables/​rbpf.tex Paired RBPF results.
papers/​anom_combined/​tables/​phone_numbers.tex Smartphone diagnostic macros.
papers/​anom_combined/​tables/​numbers.tex Generated result macros.
papers/​anom_combined/​tables/​error_model.tex Sensor error summaries.
papers/​anom_combined/​tables/​dataset.tex Dataset statistics.
papers/​anom_combined/​tables/​baselines.tex Unaided filter results.
papers/​anom_combined/​tables/​aiding.tex Paired Kalman-filter comparisons.
papers/​anom_combined/​sections/​09-availability.tex Data and software availability.
papers/​anom_combined/​sections/​08-conclusion.tex Conclusions and future work.
papers/​anom_combined/​sections/​07-discussion.tex Interpretation and limitations.
papers/​anom_combined/​sections/​06-results.tex Results and diagnostics.
papers/​anom_combined/​sections/​05-experiment.tex Dataset and experimental design.
papers/​anom_combined/​sections/​04-geophysical.tex Anomaly and sensor models.
papers/​anom_combined/​sections/​03-navigation.tex Filter formulations and settings.
papers/​anom_combined/​sections/​02-corrections.tex Corrections to conference results.
papers/​anom_combined/​sections/​01-introduction.tex Motivation and contributions.
papers/​anom_combined/​sections/​00-abstract.tex Study summary.
papers/​anom_combined/​scripts/​stats.py Loads outputs and computes statistics.
papers/​anom_combined/​scripts/​phone_mechanism.py Computes smartphone diagnostics.
papers/​anom_combined/​scripts/​build_tables.py Generates tables and macros.
papers/​anom_combined/​scripts/​build_figures.py Generates analysis figures.
papers/​anom_combined/​references.bib Bibliography.
papers/​anom_combined/​README.md Layout and rebuild instructions.
papers/​anom_combined/​NOTES.md Provenance gaps and submission tasks.
papers/​anom_combined/​main.tex Manuscript assembly and front matter.
papers/​anom_combined/​IONconf-v2.cls ION formatting template.
papers/​anom_combined/​.gitignore Excludes build artifacts.

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

Comment thread papers/anom_combined/scripts/stats.py Resolved
jbrodovsky and others added 2 commits October 2, 2026 16:03
- Differentiate the bilinear cell containing each track point for the map
  gradient, rather than sampling a grid-node central difference; regenerate
  the bound figure and emit the gradients, single-update scales and noise
  targets as macros (1.44 mGal/km, 6.37 nT/km).
- Present noise/gradient as a directional single-update heuristic, not a
  localization bound or proof of the null result.
- Generate the RBPF crossing-time and reported-sigma-vs-error statistics and
  the cross-arm sensitivity counts in scripts, with the pooling defined.
- Drop the claims that the incomplete Jacobian is conservative and that an
  uninformative likelihood redistributes particles; tie error-correlation
  claims to the phone arm's heading-dependent residual, not map decorrelation.
- State that the runs declare is_enu = true while the equations use NED.
- Correct the RBPF particle counts (2,000 full GNSS, 1,000 otherwise, per
  conf/ at 2a40f1e), qualify provenance and the availability statement, and
  fix the 2025-06-14 description (197 fixes, 12.2 m median accuracy).
- Default STRAPDOWN_DATA to the checkout's own data/ directory.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds papers/mems_nav_data: a Data in Brief (elsarticle) revision of the
rejected MEMS-Nav data paper. It adds per-device and per-trajectory statistics
computed from data/raw and data/input_real, GNSS accuracy and gravity-norm
characterization, EKF/UKF/RBPF benchmark baselines, and a frank limitations
section (the reference is the phone's own GNSS). scripts/build.py regenerates
every table, figure and quoted number. Open items (Zenodo release, ethics and
consent, mounting details, CRediT, the unverified RBPF particle count) are
marked TODO in the text and listed in NOTES.md. App device identifiers are not
reproduced.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@jbrodovsky
jbrodovsky merged commit 5e4455b into main Oct 2, 2026
9 checks passed
@jbrodovsky
jbrodovsky deleted the claude/anomaly-papers-ukf-rbpf-ekf-4d98d2 branch October 2, 2026 20:40
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants