MRTKLIB v0.4.1 Release Notes¶

Release date: 2026-03-08 Type: Feature + bugfix (RTK algorithm improvements and accuracy fix)

Overview¶

v0.4.1 ports the demo5 kinematic RTK algorithm improvements into MRTKLIB and resolves a false-fix persistence defect that caused 1σ accuracy to degrade to 1.4 m in urban-canyon scenarios. The changes affect the RTK positioning engine (mrtk_rtkpos.c) only; PPP, PPP-AR, and CLAS PPP-RTK/VRS-RTK engines are unchanged.

Highlights¶

• Partial AR (PAR) — Multi-attempt LAMBDA with polynomial ratio threshold and per-satellite exclusion loop, improving fix rates in difficult environments.
• detslp_dop() / detslp_code() — Two additional cycle-slip detectors from demo5 reduce the number of epochs with wrong integer candidates.
• varerr() rewrite — Per-constellation elevation weighting and SNR-scaled noise accurately models the satellite-specific measurement quality.
• False-fix fix (Phase 4F) — Identified and reverted a lock-counter change that caused holdamb() false-fix cycles to persist indefinitely in urban canyons. nagoya_run3 1σ restored: 1.373 m → 0.135 m (baseline: 0.128 m).

Disclaimer: The configuration parameters used here have not been tuned or optimised for maximum accuracy. Results are provided for reference only and should not be taken as representative of the best achievable performance.

What Changed¶

New configuration options¶

Three new prcopt_t fields (and corresponding rnx2rtkp/rtkrcv option keys) are available. Library defaults are conservative (disabled); the benchmark configuration activates them:

Ambiguity resolution improvements¶

Partial AR — manage_amb_LAMBDA()¶

The previous single-shot LAMBDA call is replaced by a multi-attempt loop:

1. Polynomial ratio threshold — The acceptance threshold scales with the number of satellite pairs nb; fewer pairs require a higher ratio to prevent random fixes.
2. PAR satellite exclusion — If the ratio is below threshold, the satellite with the weakest contribution is temporarily excluded and LAMBDA is retried. Iteration continues until the ratio passes or the minimum satellite count (minfixsats) is reached.
3. AR filter — If a newly-locked satellite (lock count == 0) caused the ratio to drop compared to the previous epoch, it is excluded for several epochs before being re-evaluated (arfilter=1).
4. Position-variance gate — AR is skipped entirely when the filter's position variance exceeds thresar[1] (default: 0.9999 m), preventing premature integer fixing before convergence.

Doppler cycle-slip detection — detslp_dop()¶

Restored from demo5 with a clock-jump mean-removal fix. Computes the expected phase-rate change from Doppler and compares it to the actual phase difference; a discrepancy exceeding thresdop cycles/s is flagged as a cycle slip. The clock-jump removal prevents the mean Doppler error in a single epoch from triggering false slips on all satellites simultaneously.

Observation-code cycle-slip detection — detslp_code()¶

New in MRTKLIB. Flags a cycle slip whenever the signal-tracking code for a frequency changes between epochs (e.g. L1C → L1P), indicating a receiver re-acquisition that typically introduces an integer-cycle ambiguity jump.

Measurement noise model — varerr()¶

Rewritten to match demo5:

Kalman filter improvements¶

• Adaptive outlier threshold — Phase outlier rejection threshold is inflated 10× for the first epoch or immediately after a bias reset, preventing healthy satellites from being falsely rejected during filter initialisation.
• rejc<2 guard — Phase-bias reset in udbias() is now only triggered when the outlier count reaches 2, preventing a single-epoch glitch from resetting a well-converged ambiguity.
• Acceleration coupling gate — The F[pos,acc] state-transition term is disabled when position variance exceeds thresar[1], preventing dynamics-model corruption before the filter has converged.

GNSS data quality fixes¶

• GLONASS clock rejection — satpos() now discards GLONASS ephemeris entries with |taun| > 1 s; corrupted clock entries caused position jumps.
• GLONASS health check — satexclude() switched from generic svh != 0 to the ICD-specific mask (svh&9) != 0 || (svh&6) == 4; this correctly identifies L1-bad, L2-bad, and malfunction states while accepting satellites in normal operation that the generic check would incorrectly exclude.

Miscellaneous correctness fixes¶

• seph2clk() parenthesis — Iteration guard for SBAS ephemeris clock computation had mismatched parentheses; corrected (no observable effect on supported test data).
• GF slip detector — Added thresslip == 0 early-out guard; replaced early return with continue to process all satellite pairs instead of stopping at the first pair with missing data.
• SBAS AR coupling (Copilot review) — sbas1 was incorrectly derived from the GLONASS AR state; now set independently from navsys & SYS_SBS.
• detslp_dop() nf clamp (Copilot review) — nf is now clamped to NFREQ before indexing the dopdif and tt stack arrays, matching detslp_code().
• rejc[] phase-only (Copilot review) — Code-residual outliers no longer increment rejc[]; doing so could trigger spurious phase-bias resets in udbias() via the rejc >= 2 guard.

False-Fix Persistence Fix (Phase 4F)¶

This is the most impactful single change in v0.4.1.

Root cause¶

Once holdamb() fires, it applies a Kalman measurement update that constrains ambiguity covariance to VAR_HOLDAMB = 0.001 cy². Because process noise accumulates at only ~10⁻⁸ m²/epoch, these constraints are effectively permanent until a cycle slip or satellite loss. If the wrong integers are held, the filter is locked at the wrong position with no self-correction mechanism.

A Phase 4D change had gated the lock counter on nfix > 0 && fix[f] ≥ 2:

/* Phase 4D — caused false-fix persistence: */
if (rtk->ssat[sat[i]-1].lock[f]<0 ||
    (rtk->nfix>0 && rtk->ssat[sat[i]-1].fix[f]>=2))
    rtk->ssat[sat[i]-1].lock[f]++;

When a false fix broke (nfix reset to 0 on any float epoch), lock counts froze for all non-countdown satellites. At the next epoch, LAMBDA operated on the identical satellite set with the identical constrained ambiguity state — the same wrong integers were selected again, holdamb() fired again after minfix epochs, and the cycle repeated indefinitely.

Fix¶

Reverted to unconditional lock++:

/* Phase 4F — unconditional: allows satellite-set diversification between fix attempts */
rtk->ssat[sat[i]-1].lock[f]++;

With unconditional incrementing, lock counts continue to evolve during float periods. Satellites that enter or leave the AR-eligible window between attempts introduce variation in the integer search space, enabling eventual escape from the false-fix cycle.

Effect¶

RTK Benchmark Results¶

All results use city conf (nagoya.conf / tokyo.conf), --skip-epochs 60. FIX tier (Q = 4) only. Parameters are not tuned; see the disclaimer above.

Fix rate and accuracy vs v0.3.3¶

4/6 runs improved fix rate; all 6 runs maintain Phase-3-level 1σ accuracy.

Notes: - nagoya_run2 +12 pp — The largest improvement. PAR finds valid subsets on this run even when the full satellite set produces a low ratio. - tokyo_run2 1σ (0.010 → 0.020 m) — The slight increase reflects the richer satellite set that varerr() now models with per-constellation weighting; absolute accuracy remains sub-centimetre. - nagoya_run1 fix rate (29.7 → 27.4%) — Phase 4D's conditional lock had boosted this run to 38.2% at the cost of nagoya_run3 accuracy; the Phase 4F revert returns it to near-baseline, with identical 1σ.

Full result tables¶

See docs/benchmark.md for complete three-tier (FIX/FF/ALL) tables including CLAS PPP-RTK and MADOCA PPP results.

Acknowledgements¶

RTK algorithm improvements in this release are adapted from demo5 RTKLIB by rtklibexplorer, a community fork of RTKLIB maintaining kinematic RTK accuracy improvements for challenging environments.

Benchmark results use the PPC-Dataset from the Precise Positioning Challenge 2024 (高精度測位チャレンジ2024), organised by the Institute of Navigation Japan (測位航法学会):

Taro Suzuki, Chiba Institute of Technology. PPC-Dataset — GNSS/IMU driving data for precise positioning research. https://github.com/taroz/PPC-Dataset

Test Suite (57 tests, unchanged)¶

No new CTest entries in this release. All 57 tests from v0.3.3 continue to pass.