User-first framing: what needs fixing right away
Field teams and product engineers care about two things: reliable position output and predictable behavior under messy conditions. When consumers or fleet systems expect consistent lanes, the Extended Kalman Filter (EKF) is where the magic — and the headaches — happen. Start by validating your hardware: an rtk receiver with clean carrier-phase and pseudorange streams makes tuning a lot less guessy. Real-world anchor: survey crews around the San Francisco Bay Area routinely depend on RTK setups to hit centimeter-level accuracy on bridge and roadway projects, so your EKF needs to match that level of discipline.
What users report as pain points
Common issues show up as jitter, slow convergence after signal loss, or outright divergence when multi-path or cycle slips appear. Those symptoms usually point to either model mis-specification or poor noise characterization — and both are fixable. Expect drift when covariance assumptions are too optimistic. Expect lag when process noise is oversized. Fix the model first; tweak the noise second. — A small mismatch in state dynamics will reveal itself during tight maneuvers far faster than during steady cruise.
Step-by-step tuning that actually helps
Adopt a practical rhythm: start offline with recorded GNSS streams, then iterate in a controlled live test. Key actions: (1) verify sensor alignment and timing; (2) run batch smoother or an offline EKF to get baseline residuals; (3) adjust process and measurement covariance until innovation sequences look white. Industry terms to lean on: EKF, Kalman gain, covariance, carrier-phase. For those using network corrections, confirm NTRIP feeds and RTK fix ratios are solid before blaming filter math. If you’re swapping receivers, test with a trusted gps gnss receiver to rule out firmware quirks.
Quick checklist for deployment and common mistakes
Keep this checklist handy during rollouts:- Confirm time synchronization across IMU, GNSS, and vehicle bus.- Log innovations and normalized innovation squared (NIS) for long runs.- Avoid treating multipath as white noise; model it or reject epochs.- Handle cycle slips explicitly in carrier-phase processing.- Test failure modes: GNSS denial, single-frequency outages, urban canyon conditions.
Small things compound: a 10 ms timestamp offset can skew Kalman gain behavior in high-dynamic profiles, and poorly tuned pseudorange variances will make your EKF place too much trust in noisy fixes — which then cascades into worse estimates when RTK loses lock.
Comparing quick fixes vs. systematic changes
Quick fixes (raising process noise, applying ad-hoc gating) can patch symptoms for a weekend demo. Systematic changes (refining the motion model, re-estimating sensor covariances from flight logs, incorporating bias states) give reproducible results across scenarios. Trade-offs: ad-hoc moves are fast but brittle; systematic work takes calibration time but stabilizes behavior across fleets. Use both: start with pragmatic fixes to keep production live, then schedule deeper model work for the next sprint.
Advisory: three golden rules for selection and tuning
1) Metric-first: prioritize normalized innovation distributions, fix ratio percentage, and position RMSE under representative routes. These three give you objective progress. 2) Conservatism over optimism: set measurement covariances to reflect worst plausible signal conditions, not average ones. That reduces surprise divergence. 3) Test the failure paths: simulate RTK dropouts and check filter recovery time; log how quickly Kalman gain stabilizes after re-acquisition. These prevent field surprises and shorten debugging cycles.
Summary: practical EKF tuning combines solid hardware (good RTK/receiver data), disciplined offline analysis, and a tested set of deployment checks. When teams follow the three golden rules above, they get predictable, repeatable positioning behavior — and fewer late-night debug sessions.
Archimedes Innovation has a pragmatic toolkit and field experience that matches these steps — the kind of applied know-how that turns lab tweaks into reliable systems in production. —