Repository navigation
0.5.0
One behaviour change, and it only affects the ground-robot profiles. If you set platform.profile: ground_robot or ground_robot_flat, those profiles now turn on GNSS re-anchoring, which can make the published position jump after a GPS outage. Set gnss.reanchor_enabled: false to keep 0.4.x behaviour; an explicit false wins over the profile. Nothing changes for platform.profile: general, the default. vslam.frame_id was removed (it was declared and never read).
Behaviour change: re-anchoring is on in the ground-robot profiles
This is a product decision with stated costs, not a pass. Re-anchoring failed the rule a library default change has to pass (tools/ship_rule.md, written before the test) on two secondary clauses: on the seven held-out NCLT sequences it raised median error by 3.3% and 4.4% on two, and output-rate jitter by 0.126 on one. So the library default stays off.
What it does, on all twelve NCLT sequences with the corrected dataset player (offline replay, full-trajectory ATE RMSE, tools/eval_20261025_results.md):
RMSE lower on 10, higher on none, identical on the 2 where it never fires
median higher on 4 (+0.3% to +3.0%)
jitter higher on 3 (up to +0.124 output-rate path ratio)
largest move 2012-08-20, 165.364 m -> 45.669 m
Scored at GPS fix times it brings FusionCore below raw GPS on 2 of 12 sequences, both held out (2012-03-31 by 22.0%, 2012-10-28 by 5.6%). The shipped default does not beat raw GPS on any (raw GPS lower on 11, one tie). Through the ROS node on 2012-01-08 re-anchoring read 10.882 m at fix times against 10.898 m in replay on the same points, inside a pass band fixed before that run. One sequence, one run.
Each re-anchor is a deliberate position jump, reported on /fusion/debug/gnss_status (reanchored, reanchor_jump_m, reanchor_count) and in /diagnostics. Anything that differentiates the pose will see it.
Why it exists
On NCLT, 74.8% to 96.8% of squared error on the five development sequences sat in GPS blackouts or the re-convergence after them, and in that recovery window the outlier gate rejected 81% of the fixes arriving on two sequences: the filter had drifted far enough that good fixes looked like outliers. Re-anchoring watches for that state (the gate refusing most of what arrives while the recent fixes agree with each other and with wheel odometry) and moves position, never heading, onto them.
Also in this release
- Our NCLT dataset player declared the IMU's accelerometer-derived tilt as a 0.57 degree attitude. That was our choice, not a dataset flaw. It made the node fuse the accelerometer twice and lock onto a wrong position for about 800 s on 2012-01-08 (178.808 m to 57.314 m once fixed). Every NCLT number measured before 2026-10-25 used the old player.
- The node warns once when it fuses IMU roll/pitch declared tighter than 1 degree, the mistake above.
tools/bag_report.pyreads any rosbag2 and reports why an estimator is drifting, with no ground truth and no FusionCore build: IMU and wheel yaw rates that disagree, sensors on different clocks, a receiver whose declared accuracy does not match its fixes, a log with no stop, long GNSS gaps, and an IMU orientation that is only accelerometer tilt.platform.profile(general, ground_robot, ground_robot_flat) declares what the platform can observe. It zeroes the bias process noises and priors on a robot with encoders. It slows the biases rather than holding them fixed; see the CHANGELOG.messages_loston/diagnosticsand a per-minute count of consumed sensor callbacks.- Declination sign fixed on the raw-field magnetometer path, and applied on the 9-axis IMU path, which ignored it.
What is not claimed
The comparison against robot_localization on NCLT is withheld: our player fix changed its input and broke its config, and it is being re-run. There is no hardware result in this release. Full detail, with every number's pipeline and scoring points, is in CHANGELOG.md and tools/benchmark_regression.md.