Release Notes — v0.6.10¶

SPP accuracy: C/N0 weighting, IGG-III robust estimation, and TDCP¶

Release date: 2026-05-24 Type: Feature — single-point positioning (SPP) accuracy, opt-in / default-off Branch: release/v0.6.10

Overview¶

v0.6.10 modernises the single-point positioning engine (PMODE_SINGLE, mrtk post -p 0 / mrtk run with mode = "single"), which until now was a near-verbatim RTKLIB 2.4.3 snapshot least-squares solver. It adds four staged, independently-gated improvements (P1–P4 of the plan in docs/design/spp-accuracy.md):

1. C/N0 (Sigma-ε) pseudorange weighting — down-weight low-C/N0 signals.
2. IGG-III robust re-weighting — down-weight / reject pseudorange outliers.
3. Pre-robust acceptance gate — keep the chi-square gate effective under robust weighting (the piece that makes 1+2 a clean win).
4. TDCP velocity + jump rejection — mm/s-class velocity from time-differenced carrier phase, plus rejection of code position spikes that disagree with the TDCP displacement.

Every feature is TOML-gated and off by default (prcopt_default unchanged), so any existing configuration produces bit-identical results; the gains appear only when the new keys are set (see conf/benchmark/single.toml).

This is the SPP counterpart to the v0.6.7–v0.6.9 PPP/PPP-AR work on the correction axis — a substantial new capability shipped as an opt-in patch.

Major changes (#116)¶

1. C/N0 (Sigma-ε) weighting¶

varerr() (src/pos/mrtk_spp.c) adds σ² += (fact·snr_error)² · 10^(0.1·max(0, snr_max − C/N0)), identical in form to the RTK engine's varerr(). The parameters snr_max / snr_error (err[5] / err[6]) were previously zero-initialised and unreachable from any config; exposing them also makes the RTK engine's dormant SNR term configurable. Gated by snr_error > 0 (default 0 → off).

2. IGG-III robust re-weighting + 3. pre-robust gate¶

estpos() applies an IGG-III three-segment equivalent-weight function to the standardised pseudorange residuals, scaled by a MAD robust scale (σ̂ = max(1, 1.4826·median|r̃|); Rousseeuw & Croux 1993). Robust weighting alone defeats the chi-square gate (it shrinks the residuals the gate inspects), which inflates the error tail; so when robust is active, valsol() is fed the pre-robust, all-satellite residuals — the outliers the robust pass suppresses still trigger rejection of inconsistent epochs. Gated by robust = "igg3" (default "off" → bit-identical).

4. TDCP velocity, slip detection, jump rejection¶

• pntpos() stores per-satellite carrier-phase history (ssat.ph/pt) each epoch.
• spp_detslp() flags cycle slips (LLI + Doppler-vs-phase-rate consistency).
• resdop() uses the TDCP phase rate (mm/s-class) when locked and slip-free, else falls back to Doppler (preserving the Doppler-absence behaviour).
• A jump-rejection QC drops epochs whose code position change disagrees with the TDCP displacement (velocity·Δt) by more than tdcp_jump.

Gated by tdcp = true (default false → bit-identical).

Performance¶

Measured with scripts/benchmark/run_benchmark.py --mode single on the PPC-Dataset (six Nagoya/Tokyo urban-driving runs), mean across runs, baseline (features off) → all four enabled:

The jump-rejection QC in particular collapses the catastrophic tail (e.g. one run's RMS 36.85 → 4.84 m).

Configuration¶

Enable the features under [positioning] / [kalman_filter.measurement_error], e.g. (see conf/benchmark/single.toml):

[positioning]
mode = "single"
robust = "igg3"     # IGG-III robust re-weighting + pre-robust gate
robust_k0 = 1.5
robust_k1 = 4.0
tdcp = true         # TDCP velocity + jump rejection
tdcp_jump = 5.0     # m

[slip_detection]
doppler = 1.0       # Doppler-vs-phase slip threshold (cyc/s)

[kalman_filter.measurement_error]
snr_max = 50.0      # dB-Hz
snr_error = 0.5     # m  (0 = C/N0 weighting off)

Tests¶

ctest: zero regression. The SPP regression (rnx2rtkp_spp / rnx2rtkp_spp_check) and the PPP/IGS/RTK suites pass unchanged; the two perennial failures (rtkrcv_rt — a headless-terminal environmental issue; madocalib_pppar_ion_check — a LAPACK-vs-embedded-LU numerical-noise difference) are pre-existing and unrelated. Default-off keeps all existing outputs bit-identical.

Deferred to the smartphone benchmark (#165)¶

• P5 — common-mode clock-jump correction: only relevant to receivers that step their clock (smartphones / low-cost), not the geodetic PPC set.
• P6 — position EKF (track smoothing / coasting): an untuned first cut gave no gain and diverged on the geodetic kinematic data (the post-P1–P4 WLS is already good there); it belongs on a smartphone/static benchmark where jitter is large. See docs/design/spp-accuracy.md §4.6.

A GSDC-2023 smartphone benchmark plus tuned P5/P6 re-attempt is tracked in #165.