KC-UKF research report
Report and verification suite for Compatibility-Field UKF. Strong on acute spoofing; weaker on model mismatch and chronic noise.
Research report, Monte Carlo harness, and verification ledger for KC-UKF. The JAX / KITTI navigation package is in kc-ukf-nav.
Results
CFE turns per-sensor NIS (through Gaussian kernels) into a softmax trust distribution that modulates measurement noise in the UKF update.
Scenario A — evasive UAV, acute adversarial attacks — about 7.54 m position RMSE over 100 Monte Carlo trials, roughly 20–37% better than seven baselines (Holm–Bonferroni corrected).
Where it loses:
- Ballistic re-entry: model mismatch; behind IMM-UKF
- Urban low-slow-small: chronic degradation; Sage–Husa wins
- Maritime: only after MCMC retune of CFE widths / weights — UAV defaults do not transfer
Repo contents
- Paper — XeLaTeX report (CI PDF) + CFE supplement
- Code — KC-UKF, baselines (AF/SH/Huber/WI/IMM), scenario runners, maritime MCMC, figures
- Verification — claim ledger, SymPy audit, claim-to-code pairing, rerun logs
KITTI GNSS/INS writeup: kc-ukf-nav. Archival record: Zenodo.