Question: where does the SLAM front end lose accuracy on handheld outdoor MID-360 data, and does a configuration-only change fix it?
Data:Hard Point Cloud Localization Dataset (CC BY 4.0), outdoor_hard_02a, 02b and 01b.
- IMU acceleration is rescaled from g to m/s² (tools/readme_media/scale_imu_bag.py).
- Pipeline: scripts/run_rko_lio_graph_benchmark.sh, with lidarslam/param/lidarslam.yaml for the backend.
- Metrics: APE after Umeyama SE(3) alignment against the dataset ground truth (nearest pose within 0.05 s).
- No sequence produced a loop closure, so raw and corrected trajectories coincide. The numbers are front-end (RKO-LIO) odometry.
Symptom: the 01b heading error against the ground truth jumps from about 3° to about −65° within 1.3 s, 26 s into the run, and stays there. Short-window displacement magnitudes still match the ground truth, so the failure is heading, not translation.
Cause:
The bag drops LiDAR scans for 1.2 s at that moment, and for 1.8 s later, while the IMU keeps streaming.
Past max_scan_delta_sec (default 1.0 s), RKO-LIO drops the scan and re-anchors at the new timestamp. The re-anchor keeps the pre-gap pose and discards the rotation integrated during the gap.
The operator was turning, so the next scans register against the map from a stale heading.
02a also re-anchors once, at a 1.10 s gap.
Fix:max_scan_delta_sec: 3.0. Scans after a gap of up to 3 s register from the IMU prediction instead. No re-anchor happens on any of the three sequences.
Sequence
Profile without the change
With max_scan_delta_sec: 3.0
02a
0.57 m (xy 0.46, z 0.34)
0.57 m (xy 0.46, z 0.34)
02b
0.37 m (xy 0.28, z 0.23)
0.37 m (xy 0.28, z 0.23)
01b
4.81 m (xy 4.80, z 0.31)
0.32 m (xy 0.28, z 0.16)
10 s gives the same 01b result, since the longest gap is 1.8 s.
RKO-LIO now re-anchors at the IMU-propagated pose instead of the pre-gap pose (rko_lio #16). With that fix, 01b reaches 0.32 m (xy 0.29, z 0.13) with the default max_scan_delta_sec of 1.0 s and re-levelling on, so the setting is no longer needed to avoid the heading failure.
3 s keeps the drop for genuine long dropouts, where an IMU-only prediction is no longer trustworthy.
Symptom: with the profile above, the orientation is still tilted by about 1° against the ground truth (RMS 0.7–1.0° on all three sequences), and that tilt drives most of the vertical error.
Cause:
The window measurement is right: on 02a the window tilt (1.4° at the end) matches the error against the ground truth.
A small correction is applied after every scan (02a: 2745 of 2745 scans).
The correction rotates only the new pose. The local map keeps the earlier tilt, so the next registrations pull it back.
outdoor_kidnap_a (204 s) and _b (349 s): the operator covers the sensor and carries it elsewhere.
With this profile:_a loses track 80 s into the run. _b runs to the end with 20.0 m APE.
Other settings tried: nine, covering the previous outdoor settings, voxel_size 0.5/1.0, min_icp_keypoints 10–100, no max_scan_delta_sec, and carrying only the IMU rotation across long gaps.
_a: no setting reached the end. Every run lost track 28–80 s in.
_b: every run either lost track 30–34 s in, or ran to the end with 20–34 m APE.
Truncated runs mislead: a run that loses track stops publishing poses, so its APE covers only the part before the sensor was covered. Such runs read 0.06–0.47 m. Read APE together with the trajectory duration.
Not an odometry problem: odometry alone cannot recover a pose after the sensor is carried while covered. That needs global relocalization; see configs/mid360_robot/rko_lio_mid360_kidnap_tolerant.yaml.
Source: GLIM v1.2.2 (GICP, container glim-ros2:jazzy-v1.2.2) runs from 2026-07-16, full sequences, 4 runs each.
Evaluation: the GLIM pose output was evaluated with the same alignment and association as above.
These runs were not repeated for this note, and the runtime configuration differs from ours. Read the comparison as indicative, not as a benchmark claim.
Sequence
GLIM 3D RMSE (4 runs)
GLIM xy / z (run 1)
Profile, xy / z / 3D
02a
0.72–0.80 m
0.69 / 0.21 m
0.49 / 0.24 / 0.55 m
02b
0.37–0.40 m
0.34 / 0.17 m
0.28 / 0.19 / 0.34 m
01b
0.30–0.37 m
–
0.28 / 0.13 / 0.31 m
01a
not run
–
0.65 / 0.31 / 0.72 m
GLIM still has slightly lower vertical error on 02a and 02b.