Skip to content

SOTA v7 Plan (2026-10)

This plan picks up from the SOTA Product Development Plan and the SOTA-v6 recovery plan. Their measurement policy, the fresh-holdout rules, and the SOTA-v6 runtime and memory gates still apply. What changes is the order of the work: the next bottleneck is generalisation (one configuration across datasets), not any single algorithm.

1. Where we stand (2026-10-26)

Result Evidence Why it is not yet a SOTA claim
HILTI 2022 exp04 vs GLIM CPU APE 0.048 vs 0.086 m, RTF 0.194 vs 0.218, peak RSS 369 vs 696 MB (three alternated runs, same VM) One sequence, one rival; FAST-LIVO2 (historical 0.053 m all-seven) is comparable
MapsHD humanoid MID-360, 10-sequence set Geometric mean 0.091 → 0.036 m over report chapters 15–25 Every setting was chosen on these sequences; the MapsHD frontend profile makes Oxford worse
ENWIDE (10 sequences) Lowest median ATE of five methods (0.29 m), no failure Three of ten sequences were used to build the configuration
GEODE Metro Tunneling (13 held-out bags) No failure BIEVR-LIO and FAST-LIO2 are more accurate (0.35–0.36 vs 0.84 m on tunnel1–3)
SOTA-v5 blind evaluation Geometric mean APE 24.99 m vs Voxel-SLAM 2.72 m; peak RSS 3.1× Point-LIO Failed

Every recent win depends on a configuration chosen on the same data, and the only blind evaluation failed. The fresh-holdout acquisition of 2026-08 also stalled: M2DGR's transfer endpoint returned 403, and UrbanLoco and MulRan need a request form.

2. What a claim will mean

One global configuration (G). A single frontend and backend parameter set for every sequence. Per-rig inputs are allowed: topics, frames, extrinsics, IMU acceleration unit, calibrated clock offsets, and the LiDAR beam model. The gravity-window alignment is also a per-rig input (decided 2026-10-26, report chapter 30.9). It is on only for a rig whose IMU calibration (for example static multi-orientation data, never a reference trajectory) shows a small horizontal accelerometer bias. It is off for public data without such a calibration. The tilt-correction limit with start-window gyro bias tracking is a per-rig input too (decided 2026-11-09, report chapter 60.9, G21). It is on only for a rig whose gyro datasheet publishes an angle random walk of at most 0.2 deg/sqrt(h) and an in-run bias instability of at most 1 deg/h (tactical grade, such as the STIM320), and off otherwise. Nothing else may differ by dataset or sequence. Self-adaptation inside the algorithm is allowed when it uses only sensor data and its rule and thresholds are frozen before any holdout is opened.

Claim tiers. Each tier comes from the same G and the same blind run, scored separately:

Tier Platforms Primary rivals
C0 overall every holdout sequence all pinned rivals
C1 walking and handheld humanoid, quadruped, handheld GLIM, FAST-LIO2, Point-LIO, BIEVR-LIO
C2 vehicle UGV, car GLIM, FAST-LIO2, Point-LIO, Voxel-SLAM
C3 degenerate tunnels, long corridors, open fields BIEVR-LIO, COIN-LIO, FAST-LIO2
C4 external public hidden-ground-truth leaderboards the leaderboard

Win condition for C0–C3, unchanged from SOTA-v6: lowest geometric-mean APE among the pinned rivals; no sequence more than 10% worse than its best rival; 100% completion; the SOTA-v6 runtime and memory gates. Rivals run their own published configuration with only the sensor settings changed, at revisions pinned before the run. The upstream PRBonn RKO-LIO is added as a rival, so that the claim also measures this repository's changes.

3. Data

Development data (exposed; may be tuned on)

MapsHD (8 trials), Oxford Spires Keble 05, HILTI 2022 exp01/exp04/exp07, NTU VIRAL, MID-360 RTK-SLAM, Koide handheld outdoor, ENWIDE (10), GEODE (Hard, stairs, Metro), NavINST, UrbanNav, KITTI 00/05/07. Together with the families that the SOTA-v6 recovery plan marks as consumed, none of these may appear in a later blind claim.

Holdouts (unopened)

Slot Family Tier Status (2026-10-26)
H1 UrbanLoco C2 Not available: the request form requires an institutional email address
H2 MulRan C2 Not available: the request form requires an institutional email address
H3 M2DGR C2 Blocked: the official transfer endpoint returned 403
H4 HILTI 2022 exp14, exp16, exp18 C0 (same family) Opened 2026-11-08 after the first blind run (report chapter 57); now development data. Before that: frozen unopened since 2026-08-21 (selection); a within-family holdout, weaker than a fresh family
H5 GrandTour (ETH, ANYmal legged robot) C1, possibly C3 Candidate (2026-10-26). Public on Hugging Face (leggedrobotics/grand_tour_dataset, MIT, not gated), about 50 missions, absent from this repository's records. The Livox MID-360 and Hesai XT-32 point clouds and several IMUs are per-topic files of 0.5–1.7 GB per mission. Reference trajectories are RTK-GNSS/INS (cpt7_*) and a Leica total station (prism_position). The reference topics and the dlio_* outputs of another SLAM system stay unread until the outputs are sealed. No algorithm is run on it before G is frozen. In use for C4 (2026-10-26): its COMFORT Localization Benchmark hides the references of six test missions. G was preregistered for it before any test mission was run (bench_report/data/grandtour/preregistration_G_comfort.txt); the two validation missions (HEAP-1, ARC-3) were used only to check the frame chain and are no longer holdouts. Fourteen train missions became development data on 2026-10-27 by the maintainer's decisions (bench_report/data/grandtour/dev_missions_2026-10-27.txt, dev_missions_2026-10-27b.txt). The other missions' public references stay unread, so they remain a C1 holdout. 2026-11-08: the first blind run opened 11 of them (one per site, report chapter 57); they are now development data, and 16 train missions remain unread
H6+ to be audited C2, C3 Vehicle and degenerate tiers still need families that can be downloaded without an institutional request

A tier is not claimed until it has at least two fresh families.

Cloud-VM budget

The work runs on one 4-vCPU, 15 GiB cloud VM with about 30 GB of free disk. Datasets are fetched one at a time, converted, run, and deleted. Only trajectories, logs, and hashes are kept. Ground truth is moved to a sealed directory on download and opened only after the outputs are hashed, as in report chapter 15.

4. Phases

Phase 1 — Collapse the per-dataset profiles into G (now)

The profile inventory shows 36 algorithm keys that differ from the RKO-LIO defaults across 15 representative profiles. They fall into five groups:

  1. Competitive core, shared by 5–12 families: voxel_size: 0.25, min_range: 0.7, initialization_phase: true, max_iterations: 50, convergence_criterion: 3e-5, icp_keypoint_voxel_multiplier: 2.25. This is the starting point for G.
  2. Gravity and attitude: max_expected_jerk: 1000 (4 families), min_beta: 1e9 (4, GEODE), initialization_max_gyro_rad_s: 0.02 (7), gravity_window_alignment with map relevelling (4), max_tilt_correction_deg (1). These five settings all control how much the accelerometer is trusted to correct tilt. The goal is one rule driven by measured specific-force consistency instead of five switches.
  3. Degeneracy and robustness: bump_image_registration, velocity_window_sec, skip_registration_after_gap_sec, icp_levenberg_marquardt (4–5 families, all ENWIDE/GEODE), vehicle_stall_guard (3), photometric (1). These are gated features. The test is that they stay inactive (byte-identical, or under 1% APE change) on non-degenerate data.
  4. Walking: the gait_* settings (GEODE handheld only). They must become self-gated by a GT-blind gait detector before they can enter G.
  5. Structure-specific: localizability_normal_only with max_correspondence_distance: 0.25 (MapsHD and the walking profile). They help MapsHD and hurt Oxford. The fixed value has to be replaced by one derived from the scan, for example from the point spacing or the motion-model residual.

Procedure:

  • Build G step by step (core → gravity → gated features → walking → structure). Each step runs frontend-only on every development family once (n=1).
  • Choose every self-adaptation threshold by leaving one dataset family out: choose on the others and check on the one left out. Overfitting must show up there before a holdout is spent.
  • Gate: with G, every development family is within 1.10× of its own best profile's APE, and the geometric mean is within 1.05×.

Status (2026-10-26, report chapter 30):

  • The first candidate was the shared profile plus a median-scaled Geman-McClure kernel. It roughly halves the error of the indoor MapsHD trials. It was rejected because it loses the vertical constraint outdoors: MapsHD outdoor-urk was 1.43× its best profile, and Oxford 1.13×.
  • Four other variants were also rejected.
  • Separate kernel scales for horizontal and other surfaces were rejected too: MapsHD outdoor-urk was 1.32× its best profile, and garage 1 1.14×.
  • The two outdoor development sequences point in opposite directions, so more outdoor development data comes first: Koide outdoor_hard, then ENWIDE.
  • Koide outdoor_hard needs the gravity-window keys. The shared profile is 1.51× worse there. Adding only those keys (G3) matches the Koide profile on Koide but breaks the height everywhere else: Oxford 0.88 m, MapsHD calibration-room 1 0.29 m.
  • Before any gravity key enters G, a GT-free test is needed of when the window's gravity direction can be trusted.
  • Ground-truth attitudes show why G3 fails (report chapter 30.8).
  • On Oxford and MapsHD, the window tilt turns with the heading. It comes from a body-frame accelerometer bias of 0.7–1.0°, which is larger than RKO-LIO's own attitude drift.
  • Koide's bias is 0.3°.
  • Separating the bias inside the window (G4) and raising the re-level threshold to 1° (G5) both fail.
  • No GT-free test was found for when the window may be used (chapter 30.9). The gravity window is now a per-rig input, and long handheld outdoor walks like Koide's are a known weakness of G for now.
  • G adopted (2026-10-26, report chapter 30.10): the shared profile, plus the MapsHD profile's two keys (cap 0.25 m, localizability boost 10, normal only) on scans whose keypoints' median range is below 12 m.
  • A close scan that the tight cap explains poorly (under 60% of keypoints within 0.25 m) is re-registered coarse-to-fine.
  • Over the ten development sequences, G is within 1.5% of the per-dataset best profiles (geometric mean 0.0353 against 0.0348; the shared profile alone 0.0635).
  • Configuration: configs/sota_v7/rko_lio_global_G_2026-10-26.yaml.
  • Both thresholds come from development data, so a holdout must confirm them.
  • Still open:
  • Koide-style long outdoor walks (no gravity window).
  • G now includes min_icp_keypoints: 100 (chapter 30.11), adopted by the maintainer's decision despite a +1.0% rule miss on Koide 02a. It fixes Koide 01a (3.23 to 0.98 m) and leaves the ten development sequences byte-identical.
  • More development families (ENWIDE, GEODE, NTU VIRAL) to test G before a holdout is spent.
  • G12 adopted (2026-10-26, report chapter 31.3): the GrandTour validation mission ARC-3 (multi-floor building) broke G at 4.6 m. Its stairwell scans have 75–99 keypoints, under min_icp_keypoints: 100, and were not registered. G12 adds keypoint_shortfall_voxel_multiplier: 1.0: such a scan takes its keypoints again at the 0.25 m voxel before it is dropped.
  • ARC-3 drops to 0.28 m, or 0.072 m without one 16 m jump in the reference.
  • The checked development sequences (flat-room 1, Oxford, Koide ×4), HEAP-1, and the COMFORT test trajectories are byte-identical to G.
  • Configuration: configs/sota_v7/rko_lio_global_G12_2026-10-26.yaml.
  • G21 adopted (2026-11-09, report chapter 60.9, maintainer's decision): G12 plus, on a rig with a tactical-grade gyro, configs/sota_v7/rig_tactical_gyro_2026-11-09.yaml (a 0.05 deg per-scan tilt-correction limit and gyro bias tracking in the first 30 s after a static start). On GrandTour (STIM320) it lowers the HS 15 mean from 0.0421 to 0.0289 m and the LS 8 mean from 0.0401 to 0.0376 m; every Phase 1 rig keeps G12. G16-G20 were not adopted (chapters 60.2-60.9).
  • G22 adopted (2026-11-10, report chapter 61, maintainer's decision): G21 with local_map_compact_points (the local map's points as 32-bit offsets from their voxel's corner): configs/sota_v7/rko_lio_global_G22_2026-11-10.yaml. Peak RSS 22-37 % lower on the development data, every sequence within 3.4 % in ATE; adopted despite a 0.26 % miss of its rule's LS-mean clause.

Phase 2 — Frontend accuracy, driven by the remaining development error

  • The garage scale stretch (+0.09–0.11%, specific to RKO-LIO; GLIM does not show it). The remaining hypotheses are voxels filled by their first points and point-order bias within a scan. Report chapter 29 could not separate them without per-scan comparison to the TLS map. On this VM, do that comparison in chunks streamed from the archive.
  • Long vehicle sequences are the weakest part of the record (SOTA-v5 NavINST and UrbanNav). Add one long development sequence of this kind before Phase 5.

Phase 3 — Backend: loops that never hurt, and that help on long trajectories

Order set by the maintainer (2026-10-27): the MID-360 scale stretch (Phase 2) first, then robust loops (this phase), then a multi-session mode that joins several runs of one site by place recognition (Scan Context, loop verification and candidate aggregation already exist in graph_based_slam). A wrong cross-session join breaks the whole map, so the robust loop machinery comes first. GrandTour has several runs per site (ARCHE 7, Jungfraujoch 9, Construction Halter 4); whether their total-station references share one frame is still to be checked.

  • On the short development sequences, even ground-truth loops improve APE by at most 4% (report chapters 19, 24, 27). Loops matter on long trajectories.
  • Register loop edges with GICP (the existing small_gicp plugin) instead of NDT. The NDT edges are 5–24 cm off where the frontend is 1–3 cm.
  • Derive the edge information from the registration Hessian, not from fitness (chapter 19: fitness does not predict accuracy).
  • Replace the hand-set loop rejection thresholds with a robust solve (GNC or switchable constraints).
  • Status (2026-10-29, report chapters 36-38):
  • Loop registration with small_gicp's GICP at a 0.5 m correspondence distance (registration_method: SMALL_GICP).
    • It is the first setting that does not hurt: on GrandTour no mission is more than 0.8 mm worse; on MapsHD, Oxford and exp04 no sequence is more than 2.7 % worse.
    • The adopted NDT 2 m makes two garages 16-22 % worse.
    • It does not help either: its loops are no more accurate than the frontend over the same stretch.
  • DCS (Phi 1) removes the gross loops on GrandTour but makes Oxford 24 % worse.
  • Loop information shaped by the registration Hessian (loop_edge_info_from_registration) does not help.
  • Nothing is adopted into the presets.
  • Multi-session: graph_slam_multisession_join (docs/architecture/multisession-join.md) joins two sessions of one site by place recognition, with no prior.
    • It joins all four same-site directions of two GrandTour sites within 1-3 cm.
    • In the joined map, the second session is within 5 mm of its single-session ATE.
    • It refuses different-site pairs and the stress tests.
    • On 16 more sessions split in two (rooms, corridor, garages, exp04, Oxford), a symmetric room joined 20.7 m off (turned by 180 deg). Two refusals (ambiguous hypotheses, self-similar sessions) removed every wrong join, at the cost of two right joins and ten times the run time when a join passes.
    • The self-similarity check was then sped up with the same decisions on every run:
    • a per-session cache;
    • parallel single-thread registrations, so the output is the same for any number of threads;
    • a gate on hopeless self-join candidates. A Construction pair that joins now takes 821 s instead of 1,364 s, still about 11 times a join that does not pass.
    • Each session's self-similarity can be stored under a fingerprint of its keyframes and settings (multisession_cache_dir), so a stored map is checked once.
    • A new run joined to a stored map: 573 s (only the new run's self-join).
    • A re-run with both stored: 75 s, as fast as a join that does not pass.
    • The outputs are byte-identical to computing the check.
    • With more than two sessions, each session is joined to the map of the sessions joined before it, not only to the first.
    • On 18 sessions cut into 3 parts each, 10 parts join instead of 8.
    • There were no wrong joins in 72 runs.
    • After each join the map is optimized before the next one, so later sessions match against the optimized map. Oxford's join error to the loop-closed reference halves (0.128 -> 0.064 m); the final map is unchanged.

Phase 4 — Resources

The Oxford pipeline still peaks at 1.4–1.8 GiB, against the SOTA-v6 gate of 330 MB. The main item is streaming the saved map instead of holding the full cloud (chapter 16: 685 MiB for the Oxford map alone).

Status (2026-10-30, report chapter 39): - Step 1: the saved map is downsampled from the submaps in passes over z layers, without assembling it (map_save_points_per_pass). The voxels and their order are VoxelGrid's, and the values agree within 3.8e-5 m. - Oxford's backend peak: 1,358 -> 758 MiB (984 -> 339 MiB with the PCD cache). - The pipeline sum: 1,885 -> 1,270 MiB (1,500 -> 882 MiB with the PCD cache). - The pipeline's outputs differed from run to run (three trajectories in four Oxford runs, with the frontend identical): the offline node published before graph_based_slam had subscribed, and the map started one or two scans late. The launch file now makes the offline node wait by default (wait_for_backend). Repeated default runs give byte-identical outputs (exp04 and Oxford). - Step 2: the running backend keeps each submap's scan without its padding (16 -> 12 bytes a point; outputs byte-identical). Oxford's backend while the frontend runs: 583-599 -> 500-501 MiB; its peak: 705-777 -> 654-666 MiB. - Step 3, not adopted: the offline launch's LiDAR queue at 10 frames (rko_lio's default is 50). Outputs were byte-identical and the frontend about 75 MiB smaller on average on Oxford, but both runs missed the planned 379 MiB (381 and 386 MiB). - Step 4: the running backend keeps each packed scan compressed (fields in planes, bytes shuffled, zstd level 1; outputs byte-identical). Oxford's scans take 198.8 instead of 293.5 MiB. In two pairs of alternating runs on one boot, the backend while the frontend runs: 512 -> 437 and 499 -> 427 MiB; its peak: 668 -> 595 and 649 -> 566 MiB; the save 5-7 s longer. - Step 5 (investigation): the frontend's live heap at its peak on Oxford is 363 MiB: the local map 211 (its points 143, hash buckets 48), the LiDAR queue 101, the bag read-ahead 17. About 53 MiB more of its 416 MiB heap is not live data (malloc's free memory and arena slack). - Next: malloc settings (frontend and backend), what the save adds to the backend (about 130-140 MiB), the PCD cache as the default.

Phase 5 — Blind evaluation and publication

  • Preregister G, the rival revisions, the thresholds, the machine, and the run order. Run every holdout, hash-seal the outputs, then open the ground truth.
  • First blind run (2026-11-08, report chapter 57, bench_report/data/blind_v7_2026-11-08/): G12 against FAST-LIO2, BIEVR-LIO, upstream RKO-LIO and GLIM on H4 (3) and 11 H5 missions, three repetitions, sealed before the ground truth was opened. A within-family measurement (no tier claim was possible). The win condition is not met: G's geometric mean over 14 is 0.0876 m against BIEVR-LIO 0.0706, it is more than 10 % worse than its best rival on 6 of 14, and its peak RSS exceeds 330 MB on 10 of 14; it completes all 14 in real time (RTF at most 0.49), and upstream RKO-LIO fails 2 of 14.
  • Second blind run, first family (2026-11-12, report chapter 63, bench_report/data/blind_v7_run2_2026-11-10/): G22 frozen, the same four rivals, on H8 PASTEL (multirotor, 17 sequences, a fresh family), sealed before the ground truth was opened. G has the lowest geometric mean (0.0745 m against 0.0850 for upstream RKO-LIO), is never more than 10 % worse than its best rival on the 16 scored sequences, with RTF at most 0.28 and peak RSS at most 211 MB. The win condition is not met as written: on sequence 11 the reference spans more time than the bag, so no method reaches the 95 % completion rule and G completes 16 of 17. One family supports no tier claim; H6 FusionPortable V2 and H7 MCD are preregistered and wait for Google Drive's download limit on this VM.
  • Second blind run, second family (2026-11-13, report chapter 64): H9 Odyssey (car, OS1-128, 7 drives), sealed before the reference was opened. G does not meet the win condition: it is more than 10 % worse than its best rival on all seven drives (1.23 to 1.27 on six, 5.27 on Tunnel2), third by geometric mean (0.6345 m against 0.5041 for BIEVR-LIO and 0.5231 for upstream RKO-LIO), with RTF up to 0.99 and peak RSS up to 463 MB. G loses to upstream RKO-LIO on every drive here, so one of G's changes hurts on this rig; which one is the next question. Over H8 + H9 G still has the lowest geometric mean (0.1430 against 0.1477). No SOTA claim follows from the second run.
  • Odyssey diagnosis (2026-11-13, report chapter 65; H9 is development data from here): no group of G22's keys explains the loss to upstream. The fork's default voxel downsampling is the hash-sorted one upstream reverted before 0.3.0 (upstream 0.3.2's is behind legacy_voxel_downsample, off in G22); with the key the fork at upstream's defaults returns to 0.3.2 on the two drives that are not degenerate, and G22 moves -14 %, -47 % and +24 % on the three drives tested. Voxel 0.25 is what exceeds the RTF and memory limits on the 128-beam rig. Open: a candidate on upstream's sampler (and a rebase onto upstream 0.4.x) needs its own preregistered rule on all development data, with Odyssey as a development family, and fresh blind data for a claim.
  • G23 screen (2026-11-14, bench_report/data/phase1_global/screen_g23/plan.txt): G22 with upstream's sampler holds on MapsHD (0.980) but fails exp04 (1.0511, limit 1.05) and Odyssey (geometric mean 1.122 x G22's: better on the three enclosed drives, worse on the four open ones). Dropped; G stays G22. The sampler is taken up together with a rebase onto upstream 0.4.x.
  • Rebase trial (2026-11-14, report chapter 66, bench_report/data/rebase_trial_2026-11-14/plan.txt): a merge of upstream 0.4.2 conflicts in 20 files; unmodified 0.4.1 at upstream's defaults beats G22 on Odyssey by 5.9 % with a third of the memory, but with the keys it shares with G22 it is 1.25 x (exp04) and 1.8 x (MapsHD) behind indoors, as 0.3.2 is: the indoor accuracy is the fork's features at voxel 0.25, the outdoor gain is voxel 1.0 with upstream's sampler and keypoint multiplier 1.5. Open: a GT-free choice between the two operating points by scene scale, as a candidate with its own rule; the rebase itself is worth it for memory and is a feature-by-feature port.
  • Ported base (2026-11-15, report 66 addendum): no GT-free statistic separates where voxel 1.0 beats 0.25, so no selector. Upstream 0.4.1 with four ports (IMU unit, keypoint multiplier, close-scene registration, sorted sampler) is at 0.966 x G22 on MapsHD and 0.997 x on Odyssey with a third less memory, but 1.16 x on exp04 and 1.28 x on Koide; the missing pieces are not identified. Rebase paused with the patches kept.
  • Publish only through the generated claim block in the README, one row per tier, with failed tiers shown.
  • C4: KITTI Odometry (LiDAR-only track) is the canonical hidden-ground-truth LiDAR benchmark. The HILTI 2022 ground truth is now public on the official Hugging Face dataset, so it no longer works as a hidden test. Whether the HILTI 2023 leaderboard still accepts submissions has not been confirmed.
  • C4, first entry: the GrandTour COMFORT Localization Benchmark on Codabench (LiDAR + IMU category, six test missions with hidden total-station references, SE(3)-aligned ATE, real-time methods only; test phase until 2026-10-30 23:59 UTC). G is submitted once as preregistered (report chapter 31). The maintainer uploads it, because submissions need the maintainer's Codabench account.
  • C4, ranking work (maintainer's decision, 2026-10-27; the deadline no longer sets the pace, and neither of the first two submissions goes on the public leaderboard):
  • Evaluation base. The 16 development missions, with BIEVR-LIO and FAST-LIO2 run on the same Hesai XT-32 + STIM320 input (report chapter 33). BIEVR-LIO's development mean (0.0237 m over 14 missions) is close to its test score (0.022 m), so the development mean stands in for the leaderboard.
  • Hesai first, no failures. Hesai + STIM320 is better than the Livox rig on most missions, but fails at the start and transiently, and its scan-to-scan jitter is two to three times BIEVR-LIO's (bench_report/data/grandtour/screen_hs_2026-10-27/).
  • Online loop closure with a causal output: each pose as a live system would have published it (graph_slam_offline_runner, causal_trajectory_output).
  • Several missions joined by place recognition only after the organizers confirm that other missions' data may be used; until then COMFORT runs use none.

Appendix: profile inventory

Regenerate the inventory with:

python3 scripts/rko_profile_inventory.py          # Markdown matrix
python3 scripts/rko_profile_inventory.py --json   # includes rig keys

The defaults are read from the vendored RKO-LIO sources (LIO::Config, the photometric sub-configurations, and the ROS node). Rig keys (topics, frames, extrinsics, units, clock offsets, beam model, output pacing) are listed separately in the JSON output. The 2026-10-26 matrix is in RKO-LIO profile inventory.