Vitulus ROS package for BNO085 Imu.
BNO085 over UART (~port, default /dev/imu). Watches itself (timeouts,
respawn, /bno085/health); see the docstring of nodes/bno085_node.
| topic | type | |
|---|---|---|
/bno085/imu |
sensor_msgs/Imu | game rotation vector (gyro + accel, no mag), gyro, accel, 100 Hz |
/bno085/navheading |
sensor_msgs/Imu | BNO rotation vector (with mag) — useless near the robot (median error 42°) |
/bno085/mag |
sensor_msgs/MagneticField | report 0x03, "calibrated" by the BNO dynamic calibration (it drifts) |
/bno085/mag_uncal |
sensor_msgs/MagneticField | report 0x0F, uncalibrated, one message per new report (~20 Hz), tesla, frame bno_imu_link |
/bno085/imu_status |
vitulus_msgs/Imu_status | 1 Hz |
/bno085/health |
std_msgs/String JSON | latched, 1 Hz; also mag_accuracy (0–3, BNO accuracy of report 0x03), mag_uncal (on / no data / unavailable / off; no data also when 0x0F stops for ~mag_uncal_stale_s, 2 s, while the rest streams), mag_uncal_accuracy, mag_bias_uT (the BNO's own hard-iron estimate from 0x0F) |
Report 0x0F (SH-2 "Magnetic Field Uncalibrated", 16 bytes: id, seq, status
with accuracy in bits 1:0, delay, X Y Z uncalibrated, X Y Z bias, int16 LE,
Q4 µT) is optional: if the sensor does not enable it the node runs on without
it and logs one warning. ~mag_uncal:=false switches it off. The BNO dynamic
calibration is neither enabled nor disabled by the node.
Absolute heading from the magnetometer for times without RTK: the gyro (game rotation vector) carries the heading, the magnetometer corrects it slowly. With a "trust / do not trust" gate and self-calibration from the RTK heading. Navigation does not use it yet; it is published and compared with RTK.
| topic | type | |
|---|---|---|
in ~mag_topic |
sensor_msgs/MagneticField | default /bno085/mag_uncal |
in ~imu_topic |
sensor_msgs/Imu | default /bno085/imu (game rotation vector: levelling + gyro) |
in ~rtk_topic |
sensor_msgs/Imu | default /gnss_heading/navheading; a reference only with a full fix (below) |
in ~relpos_topic |
ublox_msgs/NavRELPOSNED9 | default /gnss_heading/navrelposned: carrier solution of the RTK heading |
in ~motion_topic |
nav_msgs/Odometry | default /mobile_base_controller/odom: standing or moving (twist.linear.x) |
out /bno085/mag_heading |
sensor_msgs/Imu | ~7 Hz (every second or third report of the 20 Hz magnetometer, ~out_period_s 0.1 s), only with a calibration; stamp = time of the magnetometer sample; orientation = yaw only; orientation_covariance = [1e3,0,0, 0,1e3,0, 0,0,σ²], σ ≥ ~sigma_ok_deg (4°) when the state is ok (see Sigma), else σ² = 1e3; angular velocity / acceleration covariance −1 (not provided) |
out /bno085/mag_heading/status |
std_msgs/String JSON | latched, 1 Hz (below) |
Yaw convention: the same as /gnss_heading/navheading (ENU, yaw of Euler
'ZYX' of the quaternion, radians). The calibration offset is fitted against
that topic, so it absorbs the mounting of the IMU and the magnetic declination;
the two can be compared directly.
States (status.state):
ok— calibrated and every gate passes; σ = 4° or more (see Sigma).disturbed— calibrated, but a gate fails (reasons):field(|F − Fref|/Fref ≥ 8 %, mean of 1 s),inclination(|I − Iref| ≥ 6°),gyro_std(std of mag − gyro over 5 s ≥ 10°: a field that swings),gyro_drift(mean innovation of the fusion over 10 s ≥ 6°: the magnetometer and the gyro-carried heading disagree for seconds),anchor(the correction is ≥~max_anchor_dev_deg, 10°, further from its anchor than the gyro can have drifted: a field that is off for minutes, see Sigma),warmup(fusion window below 4.5 s, or less than 4 s of data). Published with σ² = 1e3. Near the dock the field is ~24 % off:disturbedthere is right; the heading itself is then carried by the gyro from the last good correction.uncalibrated— no calibration for this magnetometer source; nothing published.stale— no magnetometer or IMU message for~stale_s(1 s).
Status JSON: state, reasons, yaw_deg, sigma_deg, dF_pct,
dIncl_deg, gyro_std_deg, gyro_drift_deg;
fusion — tau_s, tau_eff_s (effective window now), frozen (standing:
the correction is held), freeze_standing, motion
(moving / standing / unknown), note (only when the wheel odometry is
missing: runs without the freeze), axis_weighting, resets, last_reset,
anchor (rtk / start / null), anchor_age_s, anchor_dev_deg and
hold_age_s (what sigma is made of, below);
calibration (version, age, centre, semi-axes, offset, Fref, Iref, residual
p50/p90, bins, samples, axis_weights);
self_cal (samples, bins covered/needed, last fit with its reason);
rtk_compare — rtk_valid, rtk_reason (ok, or why the RTK heading is not
a reference: carrSoln 1 (not FIXED), no NAV-RELPOSNED received, yaw variance … outside …, settling, …), rtk_require_fixed; while the RTK
heading is valid: err_deg (current error), and over the last
~stats_window_s (10 min) all and ok (median/p90/p99/max of
|error|) and ok_share. This is what a field test reads.
How it works
- Vertical hard iron hz in the sensor frame, subtracted before levelling
(h = (0, 0, hz); the horizontal part stays in the 2D centre c — subtracting
it before levelling too,
~hard_xy_pre_level, made the field strength noisier on the 2026-11-08 bag and every fit was refused). hz would otherwise leak into the horizontal on a slope as hz·sin(tilt) (30 µT at 4° ≈ 2 µT ≈ 6° with H ≈ 20 µT). h comes from a linear regression (3D h, of which hz is kept) on the RTK samples: L(B − h) = c' + A·[cos ψ, sin ψ] (ψ RTK yaw). hz is visible only through tilt; with flat samples a weak prior keeps the previous hz (0 at first). - Levelling: roll/pitch of the game rotation vector; B_level = Ry(pitch) Rx(roll) (B − (0, 0, hz)).
- 2D calibration in the levelled frame: Hc = W (B_level,xy − c) (c what is left of the hard iron, W soft iron: ellipse → unit circle); heading_raw = −atan2(Hc_y, Hc_x).
- Slow fusion with the gyro: yaw = game_yaw + d + offset. The correction d
(heading_raw − game_yaw) is a recursive least-squares estimate with
exponential forgetting, time constant
~fuse_tau_s(60 s; the game rotation vector drifts only 0.12–0.19°/min). The information starts at zero, so after a start or a restart the window grows with the data (effective constant ≈ min(elapsed, tau),fusion.tau_eff_s) and the output is usable after~fuse_init_s(5 s).-
Axis weighting. Each sample pulls d by its residual in the unit plane, Hc − (cos a, −sin a) with a = game_yaw + d, weighted by the inverse of the 2×2 residual covariance of the calibration (body axes). The field of the drive current lies along one body axis (about −145° from sensor x: 1.5 µT of residual there, 0.6 µT across), so that axis counts less. The matrix is estimated by the self-calibration from the same RTK samples (1 s means of the residual) and stored in the file (
resid_cov); a missing, non-finite or degenerate matrix means both axes count the same.~axis_weighting:=falseswitches it off. -
Standing. With the holding torque the magnetometer is off by −5…+7° while the robot stands. Standing = mean |v| of the wheel odometry < 0.05 m/s and mean |gyro z| < 0.1 rad/s over 1 s: the correction is frozen and the gyro carries the heading (
~freeze_standing, default true). The frozen correction is held unchanged for~freeze_max_s(30 min: the gyro drifts 3.6–5.7° in that time, about what the standing magnetometer is off); then its window is cut to half of~fuse_init_sand topped up from the standing magnetometer, so a robot that never moves still has a bounded heading (σ = 8° from then on). A robot that starts standing takes its first~fuse_init_sfrom the magnetometer the same way. Without the odometry (older than 0.5 s) there is no freeze and the status says so. -
Bad field. Samples that fail the
field/inclinationgate (dock, metal) are not fused. -
Game yaw origin. The origin of the game yaw moves when the IMU node restarts. A gap in the IMU data >
~imu_gap_reset_s(1 s), a yaw step faster than~max_yaw_rate(8 rad/s), or used samples more than~fuse_reset_deg(30°) from the gyro for~fuse_reset_s(3 s) start the filter again (fusion.resets,last_reset). -
Time of the sample. The game yaw is taken at the stamp of the magnetometer sample. A sample that arrives before the IMU message of its time waits for it (the two topics are not delivered in step: on the bags up to 0.13 s, which is 8° in a turn).
-
Sigma and the anchor. The gate cannot see a field that is off but inside the limits (3–5 µT in the horizontal is 8–12° and passes): the correction follows it within a minute or two, for as long as it lasts (a long fence, a steel hall). But the correction (magnetometer − gyro) can truly change only as fast as the gyro drifts, about 0.2°/min. So it is compared with an anchor:
- while the RTK heading is a valid reference (full fix, variance band, agrees with the gyro, settled) the anchor is the correction RTK implies (RTK − gyro over the last 30 messages, minus the calibration offset) and is renewed all the time; when RTK goes, it stays;
- before any RTK the anchor is the correction itself at the moment the
window first reached
~sigma_settle_s(30 s). That one is not confirmed — the filter may have started in a disturbed place — and σ stays at~sigma_start_deg(8°) until RTK has been seen; - a restart of the filter (game yaw origin) or a new calibration voids it.
deviation = |correction − anchor| (
fusion.anchor_dev_deg), andσ = max(base,
~sigma_k(0.6) × (deviation + |mean innovation over 10 s|)~gyro_drift_deg_min(0.2°/min) × time since the last fused sample)
with base =
~sigma_ok_deg(4°) when the anchor is confirmed by RTK and the window is ≥ 30 s, else~sigma_start_deg(8°). The part of the deviation that the gyro drift cannot explain (deviation − 0.2°/min × anchor age) at or above~max_anchor_dev_deg(10°) makes the statedisturbed, reasonanchor. A calibration that is 1–3° off between drives is a deviation of 1–3° and does not lift σ (0.6 × 3° < 4°). With time since the last RTK the gyro drift explains more (after an hour 12°): the state then staysok, but σ still takes the whole deviation, because a drift of the gyro and a field that crept off cannot be told apart any more. While the robot stands (correction frozen) σ grows by the full 0.2°/min of the hold.
-
- Gate as above, F = hypot(|Hc|·H0, Bz), I = atan2(−Bz, |Hc|·H0), H0 = mean semi-axis.
- Self-calibration: while RTK is valid (and has been for
~rtk_settle_s, 5 s; RTK samples that disagree with the gyro by > 15° are ignored, see below), levelled samples are taken at 5 Hz into 36 bins by RTK yaw (40 per bin, oldest replaced). Every~fit_interval_s(45 s), when ≥~min_bins(30) bins are covered: ellipse fit on the per-bin medians → drop samples with an odd field strength (> 12 %) → refit, offset → drop samples > 20° off RTK → final fit (the hard iron is re-estimated in each round). Samples older than~sample_max_age_s(7 days) are not used. Accepted when the 1 s-smoothed residual against RTK has p90 <~fit_max_p90_deg(8°), at most 30 % of the samples were dropped, it is not worse than the current calibration by more than~fit_max_worse_deg(1°) on the same samples, and not worse than the best residual so far (best_p90_deg, kept in the file) by more than that either — so the calibration cannot creep worse 1° at a time. (The best-so-far may come from a smaller, earlier sample set; a later fit on more samples is then refused until it comes within 1° — the current calibration stays in use meanwhile.) When the current calibration no longer fits the samples (p90 ≥ 8°, e.g. the robot was changed) the best-so-far is forgotten and any calibration that passes the other tests is taken. Accepted calibrations are written atomically to~calibration_file(default~/.ros/vitulus_imu/mag_calibration.yaml) and loaded at start. The file records the magnetometersource(uncal/cal): a calibration of another source is not used. - Refinement without full coverage (
~refine_calibration, on; off = nothing happens until 30 bins are covered, as before). A robot that drives dock → point → dock never heads some ways (the 2026-11-15 nights: 27 of 36 bins over six drives, −120…−30° never), so the full fit above never comes. When a fit is refused for coverage, the current calibration (same source) is checked on the same samples instead: samples with an odd field strength (> 12 % offFref) or > 20° off RTK are dropped (at most 30 %), and it must still fit them (residual p90 <~fit_max_p90_deg, 8°), with ≥~refine_min_samples(100, 20 s of valid RTK) samples in ≥~refine_min_bins(18, half the circle) bins of ≥ 3 samples. Then:- the residual covariance (axis weighting) is added to it when it has none, or refreshed when the one it has came from a refinement on fewer than 2/3 of the samples now at hand (a matrix from a full fit stays);
- only with
~refine_offset(default off) the offset gets a step = − median over the bins of the per-bin median residual (each direction counts once), only when that is clear: the bins agree (std ≤~refine_max_spread_deg, 3°), the centre is ≥~refine_min_step_deg(0.3°) and ≥ 3 standard errors, and the residual p90 is not worse with the step. One step is at most~refine_max_step_deg(2°), the sum of the steps since the last full fit at most ±~refine_max_total_deg(5°); then it waits for a full fit. σ of everyokoutput is raised by |sum of the steps|. The ellipse and the hard iron are never changed by a refinement. The reference is the same as for the full fit (full fix, variance band, agrees with the gyro, settled), so a float or glitching RTK gives no samples; repeated fits on the same samples do not add up (each one is measured against the offset as it is now). The file (version 2) recordsrefine: {count, offset_sum_deg, last, cov_samples}; a full fit starts it again. At the moment of an offset step the output does not jump: the correction and its anchor take the opposite step, so the heading moves to the new offset over the fusion window (fuse_tau_s, ~1 min; while standing frozen, until it drives); the deviation from the anchor does not change at that moment, σ grows by the step. A reference that is off by a steady amount (an RTK antenna moved, a wrong baseline for drive after drive) passes all the gates. The matrix only weights the two body axes and carries no direction, so it does not follow such an error. The offset step would: with RTK 6° off, steps of +2, +2, +1° up to the 5° limit (error inok6–7°), and after the reference is right again it comes back only a step per fit and only as far as the evidence is clear (in a check, +2° were left after two wrong and three right drives). On the drives so far the step was never needed (the axis weights removed the steady +1.5°), hence off by default. Switching~refine_calibrationoff does not undo what was written (the matrix and an offset step stay in the yaml): back = copymag_calibration.v1.yamlover it, or~axis_weighting:=falsefor the matrix. On the 2026-11-15 drives the axis weights alone removed the steady +1.5° (the offset was within the noise, no step); see "Replay 2026-11-15".
- Visibility.
status.self_calhasmissing_deg(directions of the RTK yaw without enough samples, as[from, to]ranges in degrees, from > to across ±180°),refine_enabled,refinementsandoffset_refined_deg(since the last full fit),axis_weighting, andlast_fit.refine(the last refinement's numbers and reason). A fit refused for coverage is logged once per~coverage_log_s(30 min) with the missing directions and what the refinement did; every change of the calibration is logged (matrix added / refreshed, offset step).
RTK heading as reference (self-calibration and rtk_compare): only with a
full fix, as navi_transform does — the last NAV-RELPOSNED (received at most
~relpos_max_age_s, 0.3 s, before the heading) has carrSoln FIXED and
relPosHeadingValid, and the yaw variance is within ~rtk_cov_min…~rtk_cov_max
(0.00043…0.00079 rad²). On the 2026-11-08 bag the headings with a small
variance but no full fix were a median 34° off, and fixed ones outside the
band 5° (max 19°); inside the band max 5.9°. When NAV-RELPOSNED does not
arrive the RTK heading is not taken and rtk_compare.rtk_reason says so
(~rtk_require_fixed:=false falls back to the variance band alone).
RTK against the gyro: the RTK yaw minus the game yaw is a slowly moving
offset. A valid RTK message whose offset is more than ~rtk_gyro_tol_deg
(15°) from the mean of the last 30 is ignored (counted in
rtk_compare.rtk_glitches). After 10 such messages in a row the history is
dropped and the RTK is believed again. So an RTK that is wrong for more than
~3 s at 3 Hz is taken as reference; the fit drops samples > 20° off, but the
comparison statistics include them.
Calibration file format. version: 2 adds resid_cov (2×2, unit plane).
Version 1 files (no matrix) load and behave as before the axis weighting
(both axes the same). The matrix is added to it and the file rewritten
(a) by a full fit that is refused (ratchet, worse than current) on samples
the old calibration still fits, or (b) by a refinement (item 7) when the
directions are not covered — the usual case; before 2026-11-15 case (b)
did not exist and a coverage refusal returned before the matrix was
computed, so a robot that never covers 30 bins stayed without it. After
an update the matrix comes once samples in 18 directions with valid RTK
are at hand (on the 2026-11-15 drives: 12 min into the first drive). A node from before version 2 refuses a version 2 file
(calibration version 2, it starts uncalibrated and calibrates again from
RTK). For that case the version 1 file is copied to
mag_calibration.v1.yaml next to it before the first version 2 write (once;
an existing copy is never replaced; on a file system without hard links it
is written directly; if it cannot be made the node logs one warning and
saves anyway): going back to the old code = copy it over
mag_calibration.yaml. Calibrations with a non-finite or
degenerate value are refused at load; a non-finite input or result is not
published as a heading (reasons: ["invalid"]). Files without hard_uT
load with h = 0.
A new robot (no file) is uncalibrated until it has turned through ≥ 30 of
36 directions with RTK, e.g. a slow turn on the spot or two (the directions
add up over drives and restarts, the samples are on disk).
Samples on disk. The self-calibration samples (36 × 40 with their
1 s window; compressed ~0.7 kB per sample, at most ~1 MB) are written to
~samples_file (default mag_samples.npz next to the calibration; '' =
memory only) atomically, at most every ~samples_save_s (60 s) when there
are new ones, and at shutdown; the copy is taken under the lock, the
compression and the write in the timer thread, and a write error is a
warning (once per 10 min), never a stop. At start they are loaded when the
file has the same format version and source; samples older than
~sample_max_age_s (7 days) are left out, as are samples with a
non-finite value or an RTK yaw out of range; samples up to 6 h after the
current time (a clock set back) count as fresh, later ones are left out
with a warning. A broken or foreign file is logged and the collection
starts empty. So the bins add up over restarts and drives (before: every
restart of the node started from 0). status.self_cal.samples_file and
samples_note say what was loaded.
Parameters (private): mag_topic, mag_source (uncal), imu_topic,
rtk_topic, relpos_topic, motion_topic (empty: no odometry), frame_id
(bno_imu_link), calibration_file, fit_interval_s, status_period_s,
and every key of MagHeadingEstimator.DEFAULTS (fuse_tau_s,
freeze_standing, axis_weighting, max_dF_pct, max_dIncl_deg,
max_gyro_std_deg, max_gyro_drift_deg, sigma_ok_deg, sigma_k,
max_anchor_dev_deg, anchor_from_rtk, freeze_max_s, rtk_cov_min,
rtk_cov_max, rtk_require_fixed, rtk_settle_s, self_calibrate,
min_bins, fit_max_p90_deg, refine_calibration, refine_min_samples,
refine_min_bins, refine_max_step_deg, refine_max_total_deg,
refine_min_step_deg, refine_max_spread_deg, refine_offset, ...);
node only: coverage_log_s (1800).
tools/mag_replay.py BAG [--two-pass] [--cal-bag BAG2] [--mag-topic /bno085/mag --source cal]
runs the same logic over a bag (magnetometer, IMU, RTK heading, NAV-RELPOSNED,
wheel odometry) and prints the error against RTK — all outputs, ok, moving,
standing — and the share of ok. --cal-bag BAG2 is the cross test: the
calibration self-calibrated on BAG2 is frozen and BAG replayed with it.
--set KEY=VALUE changes an estimator parameter, --dump FILE.npz writes the
outputs, --samples FILE loads the self-calibration samples at the start and
writes them at the end (with --cal/--save: drives in a row as on the
robot). Bags from before report 0x0F only have /bno085/mag (picked
automatically, source cal); a bag without /gnss_heading/navrelposned
is replayed with the variance band alone (printed).
nice -n 10 python3 -m pytest test/test_mag_heading.py -q
nice -n 10 python3 -m pytest test/test_bno085_mag_uncal.py -q
nice -n 10 python3 test/test_bno085_recovery.py [--old-ref REF]
The stored calibration of the robot (version 1, from 2026-11-14, offset
−5.16°), the two night drives that have the raw magnetometer
(nav_baseline_2026-11-15_a, then nav_d1_2026-11-15_a, the samples carried
from one to the other as on the robot). |error| against the single RTK
messages (full fix, band, valid ≥ 5 s) in ok, median / p90 / max:
| drive | before (no refinement) | with refinement |
|---|---|---|
| baseline_a (weights from 721 s on) | 1.41 / 2.18 / 4.51 | the same (no ok RTK after 721 s) |
| d1 (weights from the start) | 1.77 / 2.61 / 5.12, mean +1.54 | 0.59 / 1.62 / 5.98, mean +0.04 |
| baseline_a, the refined file | 1.41 / 2.18 / 4.51 | 0.44 / 0.99 / 3.41 (in-sample) |
| magtest3 (2026-11-14, where the old file was made) | 1.09 / 2.36 / 4.75 | 0.88 / 2.05 / 4.57 |
The share of ok does not change (72 / 77 / 83 %). The refinement added
the axis weights (380 samples in 18 directions) 721 s into the first drive
and did not move the offset (centre +0.3…+0.85°, within 3 standard errors;
the directions disagree by a std of 2.2–3°): the steady +1.5° of the nights
came from the equal axes, not from the offset. The larger maximum on d1
is four RTK messages in the first 2 s of valid RTK of the drive where the error was
already −4.3° and the removed +1.5° bias makes it −6°; the share above 4°
went from 8 to 3 messages, p99 4.3 → 3.6°.
Three drives: two of 2026-11-08 (b1, b2) and one of 2026-11-14 with the mower
running (b3: turns, straight strips, 4.3 min of mowing at 3300 rpm).
|error| as median / p90 / max in degrees, in brackets the share of the time
in ok. "Cross": the calibration comes from another drive and is frozen.
The stored file of the robot was rewritten by the live node during b3, so
that row is not an independent test.
Raw — against the single RTK heading messages (full fix, variance in the band, valid for ≥ 5 s). This includes the noise of the RTK heading itself (inside the band up to 5.9°), so the maxima are an upper bound.
| test | first version (1 s window) | now |
|---|---|---|
| cross b1 → b2, all | 3.7 / 8.9 / 12.4 | 1.2 / 3.0 / 5.9 |
cross b1 → b2, ok |
2.9 / 5.1 / 9.4 (22 %) | 1.2 / 3.0 / 5.4 (64 %) |
| cross b2 → b1, all | 5.3 / 8.9 / 14.3 | 2.1 / 3.6 / 8.4 |
cross b2 → b1, ok |
5.5 / 8.6 / 9.8 (55 %) | 2.1 / 3.5 / 7.5 (80 %) |
| cross b2 → b3 (6 days later), all | 3.4 / 6.6 / 11.4 | 2.0 / 3.7 / 7.5 |
cross b2 → b3, ok |
3.9 / 7.0 / 9.7 (50 %) | 2.1 / 3.8 / 6.1 (70 %) |
| robot file → b3, all | 2.0 / 4.6 / 9.9 | 1.1 / 2.4 / 5.8 |
robot file → b3, ok |
2.0 / 4.6 / 7.7 (64 %) | 1.1 / 2.4 / 4.8 (83 %) |
In ok the error is within the published σ in 94–99.6 % of the outputs and
within 2σ in all of them (worst 1.8σ); p90 in ok is ≤ 3.8°, so
sigma_ok_deg 4° stands as a 1σ.
Smoothed — against the RTK heading carried by the gyro (20 s median of RTK − gyro, full fix only): the RTK noise is averaged out, which shows the error of the heading itself; the maxima are a lower bound.
| test | first version (1 s window) | now |
|---|---|---|
| inside a bag (b1, b2; calibration of the same bag), all | 2.0 / 5.4 / 11.9 | 0.35 / 1.6 / 2.5 |
| cross b1 ↔ b2, all | 4.6 / 8.8 / 14.3 | 1.4 / 2.6 / 3.6 |
cross b1 ↔ b2, ok |
4.6 / 8.4 / 9.5 (26–56 %) | 1.5 / 2.7 / 3.6 (66–82 %) |
| cross b1 ↔ b2, standing | 5.2 / 9.3 / 14.3 | 1.7 / 2.3 / 3.4 |
| cross b2 → b3, all | 3.3 / 6.3 / 10.1 | 1.7 / 3.6 / 4.0 |
cross b2 → b3, ok |
3.9 / 6.8 / 9.3 (51 %) | 1.8 / 3.7 / 4.0 (72 %) |
| cross b2 → b3, mower off | 3.9 / 6.8 / 9.3 | 2.0 / 3.7 / 4.0 |
| cross b2 → b3, mower on | 2.7 / 5.6 / 10.1 | 1.5 / 3.5 / 3.9 |
| cross b2 → b3, standing | 4.7 / 7.7 / 9.3 | 2.0 / 3.8 / 4.0 |
What each part gives on the cross test b1 ↔ b2 (all outputs): without the
axis weights 2.7 / 3.3 / 4.3, without the standing freeze 1.2 / 3.0 / 5.5,
with fuse_tau_s 120 s 1.3 / 2.2 / 3.4. With the filter restarted every
60 s (the worst case for the growing window) the state ok reaches 7.9°
while the window is below 30 s — hence sigma_start_deg 8°; with the window
grown the largest smoothed error in ok is 3.6–4.0° and the raw p90 3.8° —
sigma_ok_deg 4°.
Synthetic check of what the field gate lets through (steady field in the
horizontal, no RTK while it lasts; the heading follows it and is back about
80 s after the end). Every output is either disturbed or ok with the
error inside 2σ:
| field | largest error | state | σ in ok |
worst error / σ |
|---|---|---|---|---|
| 3 µT for 180 s | 7.8° | ok |
4–7.6° | 1.7 |
| 5 µT for 180 s | 12.0° | disturbed (anchor), ok at the edges |
up to 9.9° | 1.2 |
| 3 µT for 420 s | 8.1° | ok |
4–7.8° | 1.7 |
| 4 µT for 600 s | 10.3° | ok |
4–9.0° | 1.7 |
| 5 µT for 600 s | 12.3° | disturbed (anchor) 60 %, else ok |
up to 10.2° | 1.7 |
| 0 → 6 µT over 600 s, held 300 s | 14.2° | ok to 12°, then disturbed (anchor) |
up to 10.2° | 1.3 |
| 10 µT for 180 s | 18.7° | disturbed (field, anchor) |
– | 0.9 |
(anchor confirmed by RTK before; with the unconfirmed start anchor σ is 8° or
more and the worst case is 1.5σ.) On the three drives the anchor changes
neither the accuracy nor the share of ok (the reason anchor never
appears), σ in ok has a median of 4.0°, and the same holds when RTK is
taken away 60 s after it first became valid (4–11 min without RTK).
The mower motor adds no visible disturbance: the per-sample noise is the same with the knife on and off (1.2 / 1.8 µT in sensor x / y), the mean horizontal field moves by less than 0.5 µT (about 1°), the vertical by −0.6 µT.
Not done, on purpose: compensation of the wheel power (B −= K·power) — on the cross test it gave 1.25 / 3.3 / 5.3 against 1.1 / 3.3 / 5.8 without, and inside the bags it was worse (p90 4.0 against 1.4); a deviation table by heading, a map by position and a 3D ellipsoid did not hold between drives.
Limits: the cross-drive error is mostly a constant 1–3° (the ellipse centre
and offset move between drives); the drives were short (≤ 14 min, strips ≤ 2 m)
— a field that is locally off for longer than fuse_tau_s is followed; the
anchor then lifts σ and, from 10°, the state. What it cannot see: a field
that creeps off slower than the gyro drifts (0.2°/min), and a robot that has
never had RTK since the filter started (σ stays 8°). While RTK is valid the
anchor is RTK itself, so the comparison in rtk_compare and the state are
then not independent: an error ≥ 10° against RTK reads disturbed.
Drive with three full turns, straights and arcs up to 10 m from the dock:
self-calibration accepted after ~4 min (36 of 36 bins), Fref 46.8 µT,
inclination 64.4° (Earth here ~49 µT / ~66°), vertical hard-iron 14 µT.
Live error against RTK heading: median ~4°, p90 ~7°, max ~11°; in ok
max 8.6° (hence sigma_ok_deg 8°). About 2° of the median comes from RTK
heading latency (~0.1 s) during fast turns (up to 1.6 rad/s); standing
still the error at one spot was a steady +6° (local field / heading-dependent
residual), so treat the heading as a coarse prior, not as a fine reference.