Skip to content

Map Publication Gates for Airside Hygiene

Last updated: 2026-05-24

Airside map publication must combine map quality, operational approval, safety evidence, and rollout control. The goal is to prevent stale maps, over-cleaned maps, hidden FOD, and temporary assets from reaching vehicles as if they were permanent ground truth.

Release States

StateMeaningAllowed use
draftcandidate map or overlay under constructionoffline validation only
validationautomated and human checks runningreplay, simulation, non-operational vehicle
quarantinedunresolved safety or evidence issuenot deployable
canarysigned release to limited zone/cohortmonitored operation inside approved envelope
activeproduction map for approved vehicles/routesnormal dispatch
rolled_backsuperseded due to issueincident/replay only
retiredno vehicle may use itarchive and legal hold as needed

Publication Gate Table

GateEvidenceRequired approver
source provenanceraw logs, survey dates, calibration, control points, coordinate frame, source-map acceptance packagemap owner
hygiene validationdynamic rejection, static preservation, source-map geometry QA report, FOD retention, unknown/quarantine reportV&V lead
semantic integritysemantic-map manifest, taxonomy/class-order hash, source-map qa_report_id, safety-class metrics, unknown/confidence policy, Lanelet2/vector validation, route reachability, geofence, speed/no-go overlaysautonomy lead
runtime map-load contractAutoware projection, Lanelet2, pointcloud metadata, PCD cell split, and loader smoke-test evidenceautonomy lead
operational fitstand/route availability, closure/work-zone status, sponsor constraintsairport ops
safety case deltahazard impact, residual risk, FAA AGVS/test-plan trace if applicablesafety lead
deployment readinesssigned compatibility manifest, compatible vehicle/software, rollback cache, canary monitorsfleet ops
post-release reviewmonitoring window, interventions, map disagreements, FOD ticketsrelease manager

The semantic-integrity gate is evaluated against the semantic layer produced by the offline aggregated-map semantic segmentation pipeline (../../30-autonomy-stack/perception/overview/aggregated-map-semantic-segmentation.md). That pipeline's own QA gates — held-out mIoU, per-class IoU on safety-relevant classes, cross-pass consistency, seam audit, version-to-version label churn (its §13.2-13.3) — are the upstream evidence this gate consumes; per-point confidence and provenance (its §8.6, §10.6) make the layer auditable for the safety-case-delta gate.

The semantic manifest's metrics_evidence.qa_report_id must dereference to a QA bundle that includes source-map geometry quality before this gate can pass. For MapEval-style checks, require a source_map_quality block with method, metric set, config hash, reference-map hash or no-reference waiver, alignment transform, threshold policy, failure-region digest, and pass/warn/fail/waived status. If the QA report is missing or cannot be dereferenced, treat semantic metrics as provisional even when mIoU and class recall look acceptable.

The source-map acceptance package from the map-construction pipeline must be present before semantic publication. It binds the source-map manifest hash, pose-graph digest, CRS/datum, calibration package, dynamic/static-transient/FOD/artifact/unknown layer digests, MapEval or equivalent source-map quality report, projection QA when imagery is used, and any quarantined failure regions. Publication is blocked when the package is blocked or missing; accepted_with_quarantine is allowed only if every quarantined region is reflected in the semantic tiling ledger, map-hygiene layers, route restrictions, or reviewer disposition.

The manifest must also carry outputs.map_hygiene_layer_digests and metrics_evidence.map_hygiene_metrics. These bind the semantic layer to the removal governance decision: permanent static, dynamic residual, static transient, movable-static, FOD candidate, artifact, unknown/review, and reviewer-decision artifacts must all be hash-addressed. A map cannot pass publication on semantic mIoU alone if the false-permanent, false-deletion, FOD-retention, or localization-delta metrics are missing.

The hygiene-validation gate uses the canonical Airside Map Hygiene Ground Truth Protocol as its label and reviewer-disposition source, with the V&V companion defining benchmark exchange fields and acceptance outputs. Publication is blocked when the candidate map lacks a signed static/dynamic/FOD/artifact/unknown report, rejected-object layer, reviewer decision state, or quarantine disposition for safety-critical deletions.

Training-Data Export Gate

Many semantic-map releases also export back-projected single-scan labels for the perception training flywheel. Treat that export as a separate gate from map publication. A map may be acceptable for vehicle localization while still being unsafe as a training corpus if transient objects, unresolved review regions, or pose-drift boundaries are exported as supervised positives.

Export conditionAllowed training useBlocker
permanent_static with source-map acceptance, pose-quality pass, semantic confidence, and split assignmentPositive label for fixed semantic classesMissing source-map QA, low view count, or validation/test timestamp leakage
dynamic_residualDynamic-removal auxiliary target or permanent-map negativeExported as pavement, building, marking, or other static positive
movable_staticQuarantine/context layer or ignored pseudo-labelParked aircraft, GSE, cones, or barriers exported as permanent infrastructure
static_transientHard negative, review evidence, or ignored pseudo-labelStationary people or temporary objects exported as map truth
fod_candidateHazard/FOD active-learning queue after reviewer dispositionDeleted as noise without retained evidence or exported as background
artifactArtifact/noise auxiliary target or ignored pointUsed to train against true thin structures without source-quality proof
unknown_reviewActive-learning queue onlyAny automatic supervised positive

The export manifest should include semantic class, release-state label, confidence, source-frame IDs, pose-quality bucket, split ID, reviewer state, and reason code for every exported point/voxel cluster. Publication is blocked for a training-enabled bundle if these fields are absent, even when the vehicle-facing map layers themselves pass. This mirrors the training eligibility contract in ../../30-autonomy-stack/perception/overview/3d-segmentation-training-paradigms.md.

Semantic-Map Product Mode Gate

Semantic-map releases can produce multiple products from the same source run. The publication record must declare which product modes are enabled so vehicle runtime, training, monitoring, and digital-twin consumers do not inherit each other's assumptions. In the schema-backed semantic-map manifest, this is the product_modes block.

Product modeRequired evidencePrimary blocker
runtime_semantic_mapsemantic-map manifest, runtime map contract, loader smoke test, compatibility hash, rollback packageMissing loader evidence, map-hygiene layer digest, or canary plan
training_exportrelease-state label, loss-mask policy, split manifest, source-map QA, reviewer state, projection QA if RGB was usedAny non-permanent_static point exported as a static semantic positive
hygiene_monitoringreason-coded removal sidecar, rejected-object layers, TTL/zone policy, reviewer queue, monitoring schemaDeleted/quarantined points lack reason code or downstream permission
digital_twin_transferpoint-to-mesh/3DGS/BIM transfer manifest, source point digest, projection/texture evidence, reviewer acceptanceMesh/texture label cannot be traced back to accepted LiDAR source evidence
benchmark_acceptancelocked held-out split, semantic labels, release-state labels, source-map QA, class/zone stratificationPublic proxy results used as release acceptance without local held-out evidence

A bundle may pass one product mode and fail another. For example, a map can be acceptable for runtime_semantic_map while its training_export is blocked because transient or unknown-review regions have not been masked. Conversely, a candidate can be useful for benchmark_acceptance analysis while still blocked from runtime because loader compatibility or operational approval is missing.

Change-Driven Semantic Release Gate

Map maintenance creates partial releases: a changed tile, overlay, or route restriction may need to move faster than a full re-survey. The publication gate should still evaluate each product mode independently, using the change-disposition record from ../../30-autonomy-stack/localization-mapping/maps/hd-map-change-detection-maintenance.md.

Change-driven artifactMay publish toMust not publish toGate evidence
Emergency no-go or FOD/hazard overlayruntime_semantic_map, hygiene_monitoringtraining_export as background or permanent classHazard evidence, owner, expiry/review SLA, route restriction, rollback/canary monitor
Confirmed permanent geometry patchruntime_semantic_map, benchmark_acceptance after validationtraining_export until labels are regenerated against the accepted source mapSource-map QA delta, localization replay, semantic re-run, changed-footprint manifest
Confirmed semantic reclassificationRuntime/vector layer after review, benchmark hard caseOld and new labels in the same training split without lineage forkTaxonomy action, reviewer state, class metrics, label-age and split update
Movable-static or static-transient clusterhygiene_monitoring, review overlay, optional hard-negative setRuntime permanent layer or supervised static positiveRelease-state label, TTL/zone policy, rejected-object sidecar, downstream permission
Prior-only or low-confidence changeActive-learning queue, benchmark candidate onlyAny release-truth product modePrior/model version, staleness score, current-survey requirement, blocked publication status

This gate is deliberately stricter for training export than for runtime overlays. A temporary no-go overlay can be operationally correct within minutes, but it should not become supervised ground truth until the source map, semantic label, release state, and split lineage are rebuilt.

Map Hygiene Checks

CheckPass signalBlocker
dynamic object removalghost rate below zone thresholdaircraft/GSE ghosts in localization layer
static preservationno unresolved deletion of safety-critical assetseroded stand marking, curb, pole, or boundary
FOD retentionFOD-like candidates retained as hazard/review evidencesmall hazard deleted as noise
do-not-delete hazard retentionraw, rejected, semantic, reviewer, quarantine/waiver, and placed-object evidence complete for all hazard-like candidateshazard candidate deleted, waived, or class-filtered without preserved evidence and signed disposition
movable-static policytemporary assets published only as overlayscone/barrier/GSE promoted without approval
sparse LiDAR handlingweak evidence marked unknown or reviewedunobserved area marked free
localization replayNDT/ICP health neutral or improvedresidual, covariance, or recovery regression

Release Checklist

  1. Bundle point-cloud, semantic, projection, overlay, and validation artifacts atomically.
  2. Include map ID and active layer IDs in every vehicle mission log.
  3. Sign the bundle and record compatible software, sensor, calibration, model, taxonomy, telemetry schema, and map-runtime versions.
  4. Block publication if source-map acceptance, semantic provenance, taxonomy/class-order, confidence/unknown thresholds, safety-class metrics, map-hygiene layer digests, map-hygiene metrics, QA report, or reviewer disposition is missing.
  5. Block publication if unknown regions intersect route/geofence/FOD-sensitive zones without an approved ODD restriction or quarantine decision.
  6. Block publication when the active ODD includes adverse-airside conditions but the bundle lacks signed local holdout results for do-not-delete hazards, or an explicit quarantine/ODD restriction.
  7. If the bundle exports training labels, block publication unless every exported point has a release-state label and split assignment, and unless non-permanent_static labels are masked or routed as review/auxiliary targets.
  8. Confirm Autoware map loaders launch from the signed bundle and that projection, Lanelet2, pointcloud metadata, and PCD cells are mutually consistent; if dynamic map loading is enabled, replay a representative route that requests nearby cells without unhealthy diagnostics.
  9. Confirm rollback bundle availability before canary deployment.
  10. Canary by zone, route, stand, and vehicle cohort, not by percentage alone.
  11. Monitor localization, route failures, map disagreement, semantic unknown-rate drift, FOD tickets, and interventions.
  12. Promote only after the monitoring window covers relevant conditions such as shift handover, night, rain, or busy stand operations.
  13. Retire superseded bundles only after all vehicles report leaving the old version.

Operational Overrides

OverrideRule
emergency no-gofast publish allowed; post-change review within one business day
temporary work zoneowner, reason, expiry, and briefing required
FOD hazardhazard alert can block route without permanent map edit
construction changequarantine affected tile until source evidence and route checks pass
airport sponsor restrictionoverride map route availability immediately

Sources

Public research notes collected from public sources.