MRTKLIB v0.3.3 Release Notes¶

Release date: 2026-03-07 Type: Minor (benchmark tooling — no functional changes to the library)

Overview¶

v0.3.3 introduces a kinematic positioning benchmark for evaluating MRTKLIB's three positioning modes against real-world urban driving data from the PPC-Dataset (Precise Positioning Challenge 2024). Six vehicle runs recorded in Nagoya and Tokyo are covered, with sub-centimetre ground truth from an Applanix POS LV 220 GNSS/INS system.

Three positioning modes are benchmarked:

No library code, positioning algorithms, or public API changed in this release.

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¶

Kinematic Benchmark Infrastructure (scripts/benchmark/)¶

Four new Python scripts provide an end-to-end kinematic evaluation pipeline:

Key features: - Three positioning modes: --mode clas|madoca|rtk|both|all (both=clas+madoca, all=clas+madoca+rtk, default all) - L6D archive: https://sys.qzss.go.jp/archives/l6/ (CLAS) - L6E archive: https://l6msg.go.gnss.go.jp/archives/ (MADOCA, PRN auto-probe) - Result caching: skips rnx2rtkp if output file is newer than all inputs - Per-case PNG plots (--plot) via matplotlib - Full CLI with --mode, --case, --skip-epochs, --force, -v

Three-Tier Accuracy Breakdown¶

compare_ppc.py now reports three solution-quality tiers for CLAS and RTK, plus a separate PPP tier for MADOCA:

Each row now also reports nSV (mean satellite count used in that tier). TTFF is defined as the first epoch of a ≥30-consecutive Q=4 run.

Benchmark Configurations (conf/benchmark/)¶

Five configuration files cover all hardware/mode combinations:

The mode conf and city conf are layered via two -k flags; city settings take precedence. City confs supply the precise base-station marker coordinates and the ANTENNA DELTA H from each base.obs RINEX header (ant2-antdelu).

RTK triple-frequency AR: With pos1-ionoopt=off, NF(opt)=3 and all three raw frequencies (L1/L2/L5) participate in double-difference ambiguity resolution. This is effective for the PPC-Dataset baselines (13–27 km) because the rover and base are both open-sky stations with good L5 visibility.

Receiver-specific notes (CLAS/MADOCA): - Rover RINEX files carry Unknown Unknown in ANT # / TYPE → ant1-anttype = * - Septentrio mosaic-X5 has no ISB table entry → pos2-isb = off

Bug Fixes (Benchmark Tooling)¶

Documentation (docs/benchmark.md)¶

Updated benchmark guide covers: - PPC-Dataset manual download instructions - L6 archive auto-download (python download_l6.py) - Full benchmark execution walkthrough including RTK mode - Complete option reference - Three-tier metric definitions (FIX / FF / ALL / PPP) - Updated result tables for all three modes - Configuration notes explaining the layered conf approach - Known limitations (urban multipath, long RTK baselines, no ISB, etc.)

Indicative Benchmark Results¶

Results on all six PPC-Dataset runs, GNSS-only (no IMU), --skip-epochs 60. Parameters are not tuned; see the disclaimer at the top of this document.

Nagoya¶

Tokyo¶

Interpretation:

• CLAS FIX accuracy: 0.3–1.1 m RMS 2D across all runs — consistent with PPP-RTK performance under urban multipath. Fix rates of 5–27% are expected given the dense canyon environments shown in the dataset's fisheye imagery.
• RTK FIX accuracy: Most runs achieve sub-metre FIX RMS with triple-frequency AR (l1+2+3, ionoopt=off), e.g. nagoya_run3: 0.31 m, tokyo_run3: 0.29 m. tokyo_run2 FIX RMS is elevated (18 m) despite tiny 1σ/95% values — a small number of wrongly-fixed epochs at this ~27 km baseline dominate the RMS.
• RTK FF = ALL: RTK outputs Q=4 (fix) or Q=5 (float) only — no SPP fallback — so the FF and ALL tiers are identical.
• CLAS ALL RMS: The large ALL-epoch values (13–627 m) reflect Q=1 SPP fallback epochs that dominate in the deepest canyons. The FF tier (24–78%) removes these.
• MADOCA: Float-only PPP; N reflects epochs with a valid filter solution. In nagoya_run1 (heavy multipath) only 27% of epochs produce a solution. In more open environments (tokyo_run2) coverage reaches ~100%.
• These results establish a baseline for future improvements.

Note: Kinematic benchmark results are not part of the CTest regression suite (PPC-Dataset requires a manual download). Run manually with python scripts/benchmark/run_benchmark.py.

Acknowledgements¶

The kinematic benchmark relies on 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

The dataset provides six urban vehicle runs (Nagoya × 3, Tokyo × 3) with 5 Hz triple-frequency multi-GNSS observations, 100 Hz IMU data, and sub-centimetre ground truth from an Applanix POS LV 220. We gratefully acknowledge Prof. Suzuki for making this dataset publicly available.

Test Suite (unchanged at 57 tests)¶

No new CTest entries in this release. All 57 tests from v0.3.2 continue to pass.