ENWIDE TunnelD: photometric registration on LiDAR intensity (2026-10)¶

Decision¶

Use opt-in photometric registration in RKO-LIO (rko_lio PRs #21 and #22, after COIN-LIO) with the frozen configs/enwide/rko_lio_os0_photometric_v7.yaml. On held-out TunnelS, never opened during development, RKO-LIO goes from 37.5 m to 0.75 m ATE (10 m RTE 2.0 %). COIN-LIO reproduced on the same bag reaches 0.78 m (2.2 %); its paper reports 0.743 m (1.60 %). The TunnelD numbers below are development results, because the weight and the scan-gap guard were found there.

TunnelS (held out, 251.6 m)	ATE	max	10 m RTE	path (GT 252 m)
RKO-LIO, photometric off	37.54 m	88.1 m	127.9 %	407 m
RKO-LIO, v7 (#21 and #22 identical: no LiDAR gap)	0.75 m	1.3 m	2.0 %	248 m
COIN-LIO, reproduced	0.78 m	1.3 m	2.2 %	247 m
COIN-LIO, paper	0.743 m		1.60 %

Sensitivity (reported, not used to choose): photometric_scale 0.002 gives 0.74 m (2.0 %) and 0.01 gives 1.44 m (4.1 %). The terms used about 53 patches per scan on 2377 of 2379 scans.

This is still not a degenerate_lio_sota_v1 claim. The profile's other conditions are not met: the graph backend, three repetitions, the other ENWIDE environments and the hidden tunnel.

Update 2026-10-08: v8 (below) adds three open-ground options. Over all ten ENWIDE sequences it has no failure and a median ATE of 0.29 m, against 0.69 m for COIN-LIO and 0.32 m for BIEVR-LIO, which diverges on three.

Why the earlier configurations did nothing¶

All runs below use offline_node on the converted TunnelD bag. They are scored against the Leica prism track with the IMU-to-prism lever arm, SE(3) alignment without scale, and the 10 m RTE of the profile. The scorer reproduces the July numbers.

configuration	ATE	10 m RTE	path (GT 180 m)
plain RKO-LIO	22.57 m	80.3 %	293 m
preregistered v1 (degeneracy solve, persistence gate, multiscan observability, intensity profile)	22.57 m	80.3 %	293 m
kinematic velocity blend (speed caps 4-5 m/s, gates relaxed)	22.57 m	80.3 %	293 m
voxel 0.1 m	24.13 m	75.0 %	242 m
localizability weighting	21.63 m	72.4 %	262 m

v1 is byte-identical to plain because none of its features ever engages. The degeneracy candidate appears once and is never confirmed, and the intensity prior is never attempted. Per correspondence, the translation information of point-to-point ICP is the identity, so the Hessian gates see no weak axis even while the estimate slides. The velocity blend never anchors: handheld running makes the IMU-propagated velocity disagree with ICP by up to 17 m/s. From about 18 s the ICP reports half the true speed and then oscillates near the tunnel entrance.

COIN-LIO isolates the cause¶

COIN-LIO (ethz-asl/COIN-LIO, Docker, mapping_enwide.launch, real-time playback) reproduces the paper on the same bag. Turning off only its photometric terms leaves FAST-LIO2's point-to-plane IEKF, which fails like RKO-LIO:

system	ATE	10 m RTE
COIN-LIO	0.52 m	1.7 % (paper 0.487 m, 1.59 %)
COIN-LIO, `photo_scale: 0`	25.30 m	102.8 %

The photometric patches on the intensity image are what carry this sequence.

The extrinsic of v1-v6 is wrong¶

The ROS clouds are published in os_sensor. In os_enwide.json, imu_to_sensor_transform is the identity rotation with [6.253, -11.775, 7.645] mm. The cloud-to-IMU rotation is therefore the identity, as in COIN-LIO's extrinsic_R. Columns confirm it: column 0 sits at about 170° azimuth, the os_sensor convention. v1-v6 derived T_imu_lidar = Rz(pi) for os_lidar, which is not the frame the points are in. Correcting it alone does not rescue the sequence (27.76 m ATE, 96.5 % RTE, 159 m path). v1-v6 stay unchanged as preregistration records. v7 uses [0, 0, 0, 1, -0.006253, 0.011775, -0.007645].

Photometric registration in RKO-LIO¶

rko_lio PR #21 renders each organized scan as COIN-LIO's filtered intensity image (line filter, brightness normalization, masks). It maps deskewed points back to the pixels they were captured at. Patches are selected where the image gradient sees motion along translation directions that few surface normals face. Their Gauss-Newton terms are added to every ICP iteration with weight photometric_scale^2 / N_icp.

v7 with `photometric_scale`	ATE	max	10 m RTE	path
off (corrected extrinsic)	27.76 m	45.8 m	96.5 %	159 m
0.00095 (COIN-LIO's value)	18.99 m	32.1 m	35.7 %	169 m
0.002	1.94 m	4.7 m	7.3 %	181 m
0.003 (v7)	1.70 m	4.6 m	7.0 %	180 m
0.005	1.78 m	4.9 m	7.5 %	179 m
0.01	2.12 m	5.1 m	7.2 %	177 m

The result is flat over a factor of five. RKO-LIO needs about three times COIN-LIO's weight: point-to-point correspondences also hold the pose along the weak axis, which point-to-plane residuals do not. Projecting the ICP information off the weak axis gave the same 1.70 m. That pull is not what separates RKO-LIO from COIN-LIO's 0.52 m.

Other measurements: - Repeated runs are byte-identical. - About 49 patches are used per scan, in 1184 of 1188 scans. - The 119 s bag takes about 105 s with the terms and about 70 s without, startup excluded. - Peak RSS is 422 MB against 390 MB.

Where it does not help¶

NTNU Fyllingsdalen (OS0-128 at 512 columns, fog_metadata.json geometry, line filter off) finds about 21 patches per scan. Reach is 93.7 m with the terms and 106.8 m without, of about 505 m. The tunnel is lit by a continuous LED strip and is bare concrete. As the July profile probes found, it carries almost no along-axis intensity texture. Radar remains the answer there (tunnel_radar.ros.yaml).

Scan gaps (rko_lio PR #22, merged after TunnelS)¶

The 5.1 m loop-closure error of the 1.70 m run (GT 1.1 m, COIN-LIO 1.3 m) comes from one scan. Relative to COIN-LIO's yaw, the estimate jumps 13° at 22.05 s, right after a 0.28 s gap in the bag's LiDAR stream, and keeps the error. Without photometric terms the gap is harmless. rko_lio #22 skips the photometric terms on a scan that follows a gap longer than 0.15 s:

photometric_scale	#21 ATE	#22 ATE	final yaw vs COIN-LIO, #21 → #22
0.002	1.94 m	2.38 m	−14.2° → +2.8°
0.003	1.70 m	0.34 m	about −14° → 0.0°
0.01	2.12 m	2.24 m	−13.3° → −23.0°

The guard removes the jump at every weight, but the ATE gain appears only at 0.003. Both were found on TunnelD.

Held-out TunnelS evaluation (declared before opening TunnelS; done)¶

A: rko_lio #21 with configs/enwide/rko_lio_os0_photometric_v7.yaml.
B: rko_lio #22 with the same file.
Primary metric: ATE (SE(3), prism lever arm); also 10 m RTE and path length.
Sensitivity, reported but not used to choose: photometric_scale 0.002 and 0.01 for A and B.
Merge #22 only if B is not worse than A on TunnelS.

Result: A and B are identical (0.75 m), since TunnelS has no LiDAR gap and the guard never fires, so #22 was merged.

Profile record (`degenerate_lio_sota_v2`)¶

run_enwide_sota_benchmark.sh --profile degenerate_lio_sota_v2 runs RKO-LIO with the graph backend, three repetitions, input hashes checked. v2 differs from v1 only in the frozen candidate (v7 at rko_lio e9441b3).

sequence	runs complete	ATE median	10 m RTE median	RTF	peak RSS
TunnelS (held out)	3/3	0.751 m	1.98 %	1.4-2.1	418 MB
TunnelD (development)	3/3	0.336 m	1.69 %	2.3-3.0	421 MB

All three repetitions are identical on both sequences.

The first attempts failed every repetition with exit 125, for an unrelated harness reason. In no-map mode the harness rejects any map artifact, but lidarslam.yaml lets the graph write the map bundle and pose_graph.g2o on an accepted loop closure. The July runs never closed a loop. TunnelS now returns to its start and does, so the harness sets the launch's existing M6A10_BENCHMARK_NO_MAP_ARTIFACTS marker in no-map mode.

Two operational notes: - Running other heavy jobs at the same time slowed registration to about 320 ms per scan. A repetition then ended at the harness's quiescence timeout, before the offline node had written its dump. - The node survived the process-group kill and kept publishing in the benchmark's ROS domain, which contaminated the following runs.

Run the profile on an otherwise quiet machine.

Rivals on the same bags¶

BIEVR-LIO (ethz-asl/BIEVR-LIO 2306022, ROS 2 process_bag, its enwide sensor config) is the profile's newest rival. Its paper (arXiv 2604.14421) reports ENWIDE Intersection, Runway, Field and Katzensee, but not the tunnels. On the tunnels it diverges here. It tracks for about 30 s, then slides along the axis:

	TunnelS ATE	TunnelS RTE	TunnelD ATE	TunnelD RTE
RKO-LIO v7 (profile record)	0.751 m	1.98 %	0.336 m	1.69 %
COIN-LIO, reproduced	0.78 m	2.2 %	0.52 m	1.7 %
BIEVR-LIO, reproduced	483.34 m	1194 %	58.30 m	214 %

The BIEVR-LIO sensor config uses the same cloud-to-IMU extrinsic as v7: no rotation, translation [-0.00625, 0.011775, -0.007645].

All ten ENWIDE sequences¶

The other eight sequences use one offline run per system with the same scorer and the same bag, downloaded once per sequence. The tunnel rows for RKO-LIO v7 are the profile record above; the rest of the tunnel rows are the reproductions above. COIN-LIO and BIEVR-LIO reproduce their published KatzenseeD ATE (0.592 m and 0.243 m). ATE in metres, 10 m RTE in percent:

sequence	RKO-LIO v7	RKO-LIO, photometric off	COIN-LIO	BIEVR-LIO
TunnelS (held out)	0.75 / 2.0	37.5 / 128	0.78 / 2.2	483 / 1194
TunnelD	0.34 / 1.7	27.8 / 96.5	0.52 / 1.7	58.3 / 214
KatzenseeD	0.35 / 2.0	0.70 / 3.1	0.59 / 2.2	0.24 / 1.8
KatzenseeS	0.19 / 1.1	0.21 / 1.3	0.49 / 1.9	0.19 / 1.1
FieldD	7.22 / 30.2	28.6 / 77.9	0.83 / 4.4	0.17 / 1.3
FieldS	0.69 / 2.6	6.57 / 37.4	0.20 / 1.2	0.16 / 0.8
IntersectionD	4.52 / 9.4	49.0 / 95.8	1.86 / 4.1	0.39 / 1.4
IntersectionS	0.55 / 1.3	1.39 / 3.6	0.46 / 1.5	0.23 / 1.0
RunwayD	50.1 / 50.0	34.7 / 108	2.90 / 6.5	609 / 1013
RunwayS	28.9 / 41.1	40.1 / 89.0	2.80 / 4.6	0.73 / 3.1
median ATE	0.72	28.2	0.69	0.32
runs over 20 % RTE	3	7	0	3

The photometric terms improve every sequence except RunwayD, mostly by an order of magnitude.
RKO-LIO v7 is best on both tunnels, where BIEVR-LIO diverges.
BIEVR-LIO is best on the open sequences, but it diverges on three.
COIN-LIO is the only system that never fails.
RKO-LIO v7 fails on FieldD and both runways: open, flat ground with little structure. Against BIEVR-LIO, its FieldD yaw drifts about 11° between 40 s and 100 s.

No SOTA claim follows. The profile's win policy needs zero catastrophic failures.

Open ground: v8 (2026-10-08)¶

configs/enwide/rko_lio_os0_open_ground_v8.yaml is v7 plus three opt-in rko_lio options (rko_lio 8c77478):

bump_image_registration (#23): BIEVR-LIO's voxel-wise height images replace the point-to-point residual. Relief of a few centimetres (grass, asphalt) then fixes the in-plane pose. The photometric terms stay on top.
velocity_window_sec: 0.3 (#26): the velocity is the pose difference over 0.3 s instead of one scan. On RunwayD the sensor is spun at about 190 °/s. One 0.33 m correction then turned the velocity from 2 into 5 m/s, and the weak in-plane constraint let it run away to 13 m/s.
skip_registration_after_gap_sec: 0.15 (#25): the scan after a LiDAR gap takes the IMU prediction. On TunnelD the bump registration otherwise turned the post-gap scan by 6.3° (it is constrained almost nothing in translation).

How the three were found:

On the development sequences FieldD, RunwayD and TunnelD only.
The configuration was frozen before any of the other seven sequences was run (one offline run each, same scorer, same bags as the rivals).

Two hypotheses failed first and are not in v8:

IMU biases. RKO-LIO takes them from the first scan interval, and on these handheld starts that absorbs motion: up to 0.05 rad/s in the gyro. BIEVR-LIO's online estimate converges to about (-0.023, -0.017, 0.005) rad/s on every sequence. Tracking the gyro bias online, or fixing both biases at BIEVR-LIO's values, made FieldD worse (7.2 m to 14.5–30 m). The drift there comes from the registration, not from the biases.
Photometric weight on open ground. Only 12–15 patches per scan survive on grass. A higher weight helps FieldD (5.7 m at 0.01) but breaks at 0.03.

ATE in metres / 10 m RTE in percent. Development sequences are marked *:

sequence	RKO-LIO v8	RKO-LIO v7	COIN-LIO	BIEVR-LIO
TunnelS (held out)	0.71 / 2.0	0.75 / 2.0	0.78 / 2.2	483 / 1194
TunnelD *	0.32 / 1.6	0.34 / 1.7	0.52 / 1.7	58.3 / 214
KatzenseeD	0.25 / 1.6	0.35 / 2.0	0.59 / 2.2	0.24 / 1.8
KatzenseeS	0.17 / 0.9	0.19 / 1.1	0.49 / 1.9	0.19 / 1.1
FieldD *	0.24 / 1.5	7.22 / 30.2	0.83 / 4.4	0.17 / 1.3
FieldS	0.18 / 1.0	0.69 / 2.6	0.20 / 1.2	0.16 / 0.8
IntersectionD	0.38 / 1.5	4.52 / 9.4	1.86 / 4.1	0.39 / 1.4
IntersectionS	0.19 / 0.9	0.55 / 1.3	0.46 / 1.5	0.23 / 1.0
RunwayD *	2.34 / 3.1	50.1 / 50.0	2.90 / 6.5	609 / 1013
RunwayS	0.45 / 1.7	28.9 / 41.1	2.80 / 4.6	0.73 / 3.1
median ATE	0.29	0.72	0.69	0.32
runs over 20 % RTE	0	3	0	3

On the seven sequences not used for development:

v8 has the lowest ATE on five and is within 0.02 m of BIEVR-LIO on the other two (KatzenseeD, FieldS).
No run of v8 exceeds 3.1 % RTE.
COIN-LIO is the only rival without a failure, and v8 is below it on all ten.

Mean registration time per scan is 46–83 ms on the runs that had the machine to themselves. The two Katzensee runs overlapped other jobs and took 115 and 299 ms. Peak RSS is 530–630 MB. The profile run must measure the real-time factor on a quiet machine.

This is still not a degenerate_lio_sota_v1 claim. The remaining conditions are the graph backend with three repetitions per sequence, FAST-LIO2 and Point-LIO, GEODE and the hidden tunnel.

Profile record (`degenerate_lio_sota_v3`, 2026-10-08)¶

run_enwide_sota_benchmark.sh --profile degenerate_lio_sota_v3 ran v8 at rko_lio 8c77478 on all ten sequences:

Graph backend, three repetitions each, input hashes checked.
One sequence at a time, on a machine with nothing else heavy running.

sequence	runs complete	ATE median	10 m RTE median	RTF	peak RSS
TunnelS (held out)	3/3	0.708 m	2.02 %	1.29–1.32	484–627 MB
TunnelD *	3/3	0.324 m	1.62 %	1.49–1.51	480–626 MB
KatzenseeD	3/3	0.246 m	1.57 %	1.62–1.64	590–595 MB
KatzenseeS	3/3	0.172 m	0.95 %	1.31–1.33	522–530 MB
FieldD *	3/3	0.238 m	1.55 %	1.52–1.53	622–645 MB
FieldS	3/3	0.177 m	0.95 %	1.35–1.37	613–628 MB
IntersectionD	3/3	0.382 m	1.53 %	1.57–1.58	588–594 MB
IntersectionS	3/3	0.194 m	0.94 %	1.27–1.28	586–593 MB
RunwayD *	3/3	2.341 m	3.08 %	1.53–1.55	609–618 MB
RunwayS	3/3	0.450 m	1.73 %	1.31–1.33	612–618 MB

All 30 runs completed with at least 99.6 % of the ground truth matched.
The three repetitions are identical on every sequence and equal the offline runs above.
The real-time factor is 1.27–1.64, so every sequence runs faster than real time.
Median over the ten sequences: ATE 0.285 m, 10 m RTE 1.56 %.

The harness first scored the IMU trajectory without the 11 cm IMU-to-prism lever arm (fixed in #490). The numbers above re-score the same saved trajectories with it. That changed the ATE by at most 0.004 m. The runs are deterministic, so they were not repeated. The v2 record above is still the uncorrected score.

This meets the profile's execution contract on ENWIDE. A SOTA claim still needs FAST-LIO2 and Point-LIO on the same bags, GEODE and the hidden tunnel.

Profile record (`degenerate_lio_sota_v4`, v10, 2026-10-11)¶

v10 is v8 plus rko_lio's bump-registration fallback with a persistence condition:

The fallback (#28). A scan whose bump result turns the pose more than 2° from the IMU prediction can be registered again without the bump terms.
The persistence condition (#29). That happens only when at least 3 of the last 10 scans exceeded 2°.

It was developed on TunnelD, FieldD, RunwayD and three GEODE sequences. The other seven ENWIDE sequences had already been run with v9, the same fallback without the persistence condition, so this record is not a fresh validation of them. docs/research/geode-hard-evaluation-2026-10.md has the design and the GEODE results.

run_enwide_sota_benchmark.sh --profile degenerate_lio_sota_v4 ran v10 at rko_lio cce5cf4 on all ten sequences, with the same settings as v3:

Graph backend, three repetitions each, input hashes checked.
One sequence at a time.
Another project's jobs were running on the machine, which shows in the RTF but not in the deterministic trajectories.

sequence	runs complete	ATE median	10 m RTE median	RTF	peak RSS	v3 (v8) ATE
TunnelS (held out)	3/3	0.708 m	2.02 %	1.33–1.34	487–631 MB	0.708 m
TunnelD *	3/3	0.341 m	1.67 %	1.54–1.58	478–548 MB	0.324 m
KatzenseeD	3/3	0.246 m	1.57 %	1.67–2.10	590–603 MB	0.246 m
KatzenseeS	3/3	0.172 m	0.95 %	1.37	517–533 MB	0.172 m
FieldD *	3/3	0.238 m	1.55 %	1.59–1.60	623–642 MB	0.238 m
FieldS	3/3	0.177 m	0.95 %	1.40–1.42	613–625 MB	0.177 m
IntersectionD	3/3	0.382 m	1.53 %	1.62–1.64	593–603 MB	0.382 m
IntersectionS	3/3	0.194 m	0.94 %	1.31–1.32	581–589 MB	0.194 m
RunwayD *	3/3	2.341 m	3.08 %	1.61–1.62	620–626 MB	2.341 m
RunwayS	3/3	0.450 m	1.73 %	1.36–1.51	607–620 MB	0.450 m

Completeness. All 30 runs completed with at least 99.58 % of the ground truth matched. The three repetitions are identical on every sequence.
Against v3. v10 equals v8 on nine sequences. On TunnelD, a development sequence and a tunnel, it is 0.017 m worse: a few scans there exceed 2° in short runs and fall back.
Medians. Over the ten sequences, ATE is 0.293 m (v3: 0.285 m) and 10 m RTE is 1.56 % (v3: 1.56 %).
Profile rule. The profile's 2 % limit applies to non-degenerate sequences, and v10 changes none of them.

On GEODE Hard, v10 has a lower median ATE than v8 (81 against 160 m) and keeps every sequence v8 tracks. That makes it the single configuration for both datasets. (An earlier version also called this the lowest median ATE of all methods. That comparison gave the Livox device an IMU 7.5–80 ms off the LiDAR clock; see the correction in the GEODE note.)

All profile rivals (2026-10-08)¶

FAST-LIO2 (7cc4175) and Point-LIO (4b86a46), the profile's two remaining rivals, ran on the same ROS 1 bags:

Build: docker/enwide_rivals_ros1.Dockerfile.
Run: scripts/run_enwide_rival_ros1.sh.
Configuration: configs/enwide/rivals/. Each is the method's own ouster64.yaml with only the topics, 128 lines, the 0.65 m blind range of COIN-LIO's ENWIDE config and the os_sensor extrinsic changed.
The ring field of these clouds is uint16 and the methods expect uint8. Only their feature-extraction path reads it, and that path is off for Ouster.

Every method below is scored the same way:

score_position_only_trajectory.py on the IMU trajectory moved to the prism.
v8 is the profile record (median of three identical runs).
The rivals are one real-time playback each.

ATE in metres / 10 m RTE in percent:

sequence	RKO-LIO v8	COIN-LIO	BIEVR-LIO	FAST-LIO2	Point-LIO
TunnelS (held out)	0.71 / 2.0	0.78 / 2.2	483 / 1194	38.5 / 177	37.6 / 93.5
TunnelD *	0.32 / 1.6	0.52 / 1.7	58.4 / 236	27.5 / 107	30.8 / 140
KatzenseeD	0.25 / 1.6	0.59 / 2.2	0.24 / 1.8	1.29 / 3.4	0.39 / 1.8
KatzenseeS	0.17 / 0.95	0.49 / 1.9	0.19 / 1.1	1.13 / 3.5	0.32 / 1.6
FieldD *	0.24 / 1.6	0.83 / 4.4	0.17 / 1.3	17.7 / 21.9	33.2 / 111
FieldS	0.18 / 0.95	0.20 / 1.2	0.16 / 0.8	0.19 / 1.4	0.32 / 3.1
IntersectionD	0.38 / 1.5	1.86 / 4.1	0.39 / 1.4	14.3 / 14.2	48.5 / 88.1
IntersectionS	0.19 / 0.94	0.46 / 1.5	0.23 / 1.0	13.1 / 22.1	4.15 / 14.5
RunwayD *	2.34 / 3.1	2.90 / 6.5	609 / 1012	68.0 / 121	52.2 / 107
RunwayS	0.45 / 1.7	2.80 / 4.6	0.73 / 3.1	44.6 / 103	892 / 1018
median ATE	0.29	0.69	0.32	16.0	32.0
runs over 20 % RTE	0	0	3	6	6

Neither FAST-LIO2 nor Point-LIO holds the tunnels or the runways. Both fail on six sequences.
v8 has no failure and the lowest median ATE.
On the seven validation sequences v8 has the lowest ATE on five and is within 0.02 m of BIEVR-LIO on the other two.

All four profile rivals are now reproduced on identical inputs. A SOTA claim still needs GEODE's degenerate sequences and the hidden tunnel.

Next¶

GEODE's degenerate sequences and the hidden tunnel, the profile's remaining claim conditions.