Skip to content

Perception Free-Space and Unknown-Object Safety Case

Last updated: 2026-05-09

Purpose

This safety-case page defines the argument, evidence, monitors, and operational response for perception free-space and unknown-object handling. It focuses on the highest-risk failure: the autonomy stack believes space is traversable when it is occupied, unknown, out-of-distribution, stale, or insufficiently observed.

Top-Level Claim

The vehicle only plans through free space that is sufficiently observed, current, aligned to the active map and pose, and valid for the current ODD. Unknown objects, unknown cells, OOD occupancy, and stale or conflicting evidence are handled conservatively before they can create an unsafe interaction with aircraft, people, GSE, infrastructure, FOD, or protected zones.

Hazard Table

HazardCauseSafety requirementEvidence
False free space near aircraftReflection, sparse LiDAR, bad map, wrong calibrationHigh-confidence free-space false positive is a release blocker in protected zonesClosed-course fixtures, replay labels, calibration drift injection
Unknown object ignoredOpen-set object not in training ontology, low confidence, novel GSEUnknown/OOD object in route corridor is treated as obstacle or review-triggering hazardOpenAD/OOD tests, internal airside unknown-object set
Small hazard over-cleanedFOD, chock, cone, hose classified as noiseSafety-relevant small objects are retained as current hazards until clearedFOD benchmark, map-cleaning false-deletion test
Stale obstacle/free-space gridTiming fault, dropped sensor, TF future/past extrapolationStale data is rejected and triggers degraded modeTimestamp shift, latency/jitter, message-filter validation
Wrong pose makes map free space unsafeLocalization drift or wrong map tilePose integrity red blocks autonomous motionSLAM benchmark, map-localization timing gates
Map says clear but site changedTemporary GSE, parked aircraft, expired overlayPerception-map disagreement creates ticket or route holdMap hygiene monitoring and source-session evidence
ODD condition invalidates perceptionHeavy rain, wet glare, fog, de-icing spray, low sunODD monitor restricts operation or increases conservatismAdverse-condition replay, site shadow mode

Safety Requirements

IDRequirementRuntime action
FS-01Unknown occupancy is not traversable inside route corridor, aircraft clearance envelope, pedestrian zone, geofence edge, or stop regionPlanner treats unknown as blocked or reduces speed to validated crawl behavior
FS-02Free-space output includes observation age, source sensors, unknown ratio, and confidence/uncertaintyMonitor rejects stale or unsupported free-space
FS-03Unknown/OOD object score above threshold near route corridor produces an obstacle, stop, or remote-review requestDo not suppress solely on low class confidence
FS-04Free-space and object outputs are tied to valid pose, map tile, calibration, and time stateContract violation marks perception-SLAM red/unknown
FS-05Map-cleaning and dynamic-object removal never delete FOD or temporary hazards without review evidenceMap tile quarantine and safety review
FS-06Any release-changing threshold has an owner, validation partition, locked test partition, and field monitorSUMS/safety board approval
FS-07Field incidents preserve raw evidence before aggregation or retention expiryIncident trace includes bag/MCAP, active manifest, monitor actions

Argument and Evidence

ClaimEvidence neededAcceptance rule
C1 free-space semantics are conservativeInterface contract, planner integration test, occupancy encoding testUnknown/free/occupied encoding cannot be inverted or silently changed
C2 perception detects safety-relevant obstaclesLabeled object/FOD/aircraft/GSE/person slices, false-negative reviewNo high-risk class regression in locked release slice
C3 unknown and OOD are actionableOOD/unknown benchmark, monitor-to-action test, field alert routeUnknown/OOD alert leads to obstacle, stop, route hold, or review
C4 timing and localization do not invalidate free-spaceTimestamp sweep, TF/message-filter validation, pose integrity testsFree-space is rejected when pose/time are red
C5 calibration drift is detected before unsafe outputMiscalibration injection and downstream free-space impactMonitor red occurs before false-free-space threshold is exceeded
C6 maps do not overwrite current hazardsMap hygiene QA, FOD retention tests, overlay expiry testsHazard/current layer wins over stale permanent free-space
C7 safety case stays live after deploymentDashboard, incident review, CAPA trace to claimsField alerts update assumptions, thresholds, or ODD restrictions

Runtime Monitors

MonitorSignalRed condition
Free-space conservatismUnknown area, observation age, sensor support, false-free-space sentinel zonesFree cell lacks current observation in protected route zone
Unknown/OOD objectOOD voxel/object score, open-world detector, track persistence, cross-modal disagreementNovel/unknown object overlaps route corridor or aircraft clearance
Sensor-fusion consistencyLiDAR/radar/camera disagreement, track residuals, point densityRedundant modalities disagree beyond validated envelope
Pose-map integrityPose covariance/protection level, scan-match residual, map tile IDPose uncertainty exceeds route threshold or wrong tile active
Calibration healthPair residuals, projection confidence, time offsetA consumed pair exceeds red threshold
Timing healthPTP/PHC state, stamp age, skew, TF failuresStale/future data could feed planner

Release Gates

GatePass conditionBlock condition
S0 traceabilityEach requirement maps to hazard, test, monitor, and release evidenceOrphaned free-space or unknown-object requirement
S1 nominalFree-space/object metrics pass by airside ODD sliceAggregate pass hides aircraft/person/FOD slice failure
S2 open-worldUnknown/OOD objects produce conservative action in replay and closed-course testsUnknown object is confidently mapped to free space
S3 corruptionRain, beam loss, glare, timing skew, extrinsic drift increase uncertainty or trigger degradationSilent overconfidence under credible corruption
S4 closed-courseFixtures for aircraft/GSE/FOD/cones/chocks/protected zones passAny false-free-space in protected zone
S5 shadow modeField disagreement and unknown-object alert rate inside validation envelopeAlert cluster unresolved or operator intervention pattern worsens
S6 operationsRunbooks, training, incident evidence retention, and stop authority are activeNo operational response for red monitor

Operational Response

EventVehicle responseFleet response
Unknown object in route corridorSlow, stop, or remote-assist according to route riskCreate event; preserve log window; review label/ontology
Free-space monitor redControlled stop unless validated crawl mode appliesBlock canary promotion; inspect timing, calibration, map, weather
Map-perception disagreement clusterAvoid/hold affected zoneQuarantine tile or create temporary overlay
ODD boundary exceededEnforce ODD restrictionUpdate site operations and safety-case assumption
Repeated false alarmKeep safe behavior; tune only through release processAnalyze false positives with safety lead approval

Residual Risk Rules

  • Public datasets and leaderboards do not prove airside safety; they only support generic comparability and regression detection.
  • Low false-positive pressure cannot justify suppressing unknown objects near aircraft, people, geofence boundaries, or FOD-sensitive zones.
  • A monitor that only logs but does not affect behavior is diagnostic evidence, not a safety control.
  • Unknown telemetry or missing diagnostics cannot be counted as safe exposure.
  • 60-safety-validation/safety-case/safety-case-evidence-traceability.md
  • 60-safety-validation/runtime-assurance/online-perception-monitoring-odd-enforcement.md
  • 60-safety-validation/runtime-assurance/monitor-qualification-evidence.md
  • 60-safety-validation/verification-validation/uncertainty-calibration-perception-slam-release-gates.md
  • 60-safety-validation/verification-validation/robustness/fod-retention-map-cleaning-safety-case.md
  • 50-cloud-fleet/observability/map-hygiene-operational-monitoring.md

Sources

Public research notes collected from public sources.