Skip to content

Perception-SLAM Corruption and Fault Injection Protocol

Last updated: 2026-05-09

Purpose

This protocol defines the robustness campaign for perception, localization, SLAM, and map-change systems under credible airside corruptions and injected faults. The goal is not only to measure accuracy loss. The campaign must show that failures are detected, uncertainty rises, safety monitors react, map publication is blocked when needed, and fleet telemetry preserves enough evidence for root cause analysis.

This protocol feeds the perception-SLAM evidence case, uncertainty calibration release gates, SLAM map benchmark protocol, airside adverse conditions validation, and online perception monitoring and ODD enforcement.

Campaign Inputs

InputRequired content
Candidate buildSoftware hash, model weights, runtime parameters, safety monitor thresholds
Map packageTile hashes, semantic layers, coordinate frames, source traversals
Sensor calibrationIntrinsics, extrinsics, time-sync model, calibration date, residual report
Data manifestClean logs, closed-course logs, public/proxy datasets, airside scenario tags
Fault matrixCorruptions, severity levels, random seeds, expected monitor response
Release thresholdsMetric degradation limits, safety action requirements, block conditions

Corruption Matrix

CategoryFaults/corruptionsAirside relevanceRequired severity levels
LiDAR weatherRain attenuation, fog backscatter, snow/spray points, wet-ground specular lossTropical rain, de-icing spray, fog, standing waterlight, moderate, severe
LiDAR hardwareBeam dropout, channel bias, crosstalk, range noise, intensity drift, partial occlusionDamaged/dirty sensor, connector issues, lens contamination5, 15, 30, 50 percent affected
CameraMotion blur, low light, glare, lens dirt, over/under exposure, compression artifactsNight apron lighting, wet glare, depot offload compression3-5 severity steps
RadarFalse tracks, missed tracks, multipath, velocity noiseJet exhaust zones, wet apron, metallic clutternominal plus degraded modes
GNSS/INSDropout, multipath bias, heading drift, wheel slip, IMU biasTerminal multipath, indoor/depot, wet surfacebounded bias and outage duration grid
Time synchronizationSensor timestamp skew, jitter, dropped frames, reorderingPTP/NTP failure, bus congestion10 ms to safety-critical threshold
ExtrinsicsCamera/LiDAR yaw/pitch/roll shift, LiDAR/IMU offset, sensor mount vibrationMaintenance error, impact, thermal/mechanical driftsmall detectable to unsafe
Map faultsStale tile, wrong tile, shifted tile, missing layer, dynamic object promoted staticConstruction, aircraft/GSE ghosts, bad map updatetile-level and feature-level
Dynamic sceneAircraft/GSE/person/FOD injection, occlusion, temporary barriersApron operations and stand turnaroundsingle actor, dense actors, occluded
Compute/runtimeDelayed node, CPU/GPU saturation, memory pressure, dropped diagnosticsEdge compute overload during logging or adverse weatherp95 latency to timeout

Fault Injection Rules

  1. Faults must be deterministic under recorded random seeds.
  2. Clean and corrupted runs must use the same logs and build unless the test is a hardware-in-loop or closed-course injection.
  3. Injected faults must preserve physically plausible timing and coordinate frames unless the test explicitly targets malformed data handling.
  4. Sensor corruptions must affect raw or near-raw inputs where possible, not only final detections.
  5. Map faults must include publication-path tests: candidate map, quarantine state, rollback, and runtime lookup.
  6. Each fault has an expected safe response: continue, degrade, stop, quarantine, alert, or block release.

Metrics

MetricDefinitionRobustness interpretation
Corruption error ratioMetric under corruption divided by clean metricMeasures degradation relative to same scenario
Relative robustnessAverage retained performance across severitiesUsed for model comparison, not safety alone
Localization availabilityFraction of time pose is valid and within error envelopeMust degrade gracefully under faults
Silent failure rateFault cases where output is wrong and confidence/monitor does not flagRelease-blocking for high-risk slices
Detection latencyTime from injected fault to monitor alert/actionMust be within safety budget
False-free-space under faultFault cases creating traversable output where occupied/hazardousZero tolerance in protected zones
Map quarantine recallFraction of unsafe map changes blocked before publicationCritical for fleet map operations
Recovery timeTime from fault end to stable nominal operationRequired for operational availability
Evidence completenessFraction of fault runs with required logs/diagnostics/eventsRequired for root cause and safety case

Pass/Block Rules

RuleDecision
Any high-confidence false-free-space result near aircraft, people, FOD, or geofenceBlock release
Any wrong-pose condition that remains inside nominal uncertainty bounds past safety budgetBlock release
Any map fault that reaches publication without quarantine/review when it changes protected geometryBlock release
Severe corruption causes controlled stop with correct diagnostics and no unsafe motionPass with operational availability note
Moderate corruption exceeds accuracy threshold but uncertainty and degraded mode trigger correctlyPass only if ODD restriction or mitigation is approved
Fault run lacks required evidence logsInconclusive; rerun or block if rerun impossible

Test Campaign Phases

PhaseEnvironmentPurpose
R0 static analysisConfig and manifestsVerify all fault hooks, thresholds, and expected monitor actions are defined
R1 offline replayClean and corrupted logsHigh-volume deterministic comparison
R2 simulationScenario generator and digital twinExplore rare/dangerous aircraft, FOD, and geofence cases
R3 HIL/SIL timingVehicle compute and runtime middlewareValidate latency, dropped frames, overloaded nodes, timestamp faults
R4 closed courseInstrumented physical testValidate sensor contamination, GNSS denial, wet surface, FOD fixtures
R5 shadow modeReal airside routes under supervisionConfirm event rates and monitor behavior without autonomous risk

Airside Fault Scenarios

ScenarioInjectionExpected behavior
Wet stand approachGround returns removed/specular, aircraft reflection addedFree space becomes unknown or conservative; no high-confidence clearance
Aircraft absent/present map pairParked aircraft points promoted into candidate static mapMap QA flags movable-static and blocks permanent publication
De-icing adjacencySpray-like LiDAR/camera corruption and point-density dropSensor health rises, speed reduced or ODD excludes zone
Taxiway crossing GNSS multipathGNSS position bias and IMU heading driftCross-sensor residual rises, pose covariance inflates, geofence remains conservative
Depot dense clutterTemporary carts and barriers inserted/removedMap-change workflow quarantines tile or requires review
Time-sync fault during moving actorsCamera/LiDAR skew injected while person crossesFusion uncertainty rises; no stale detection used as current obstacle
Sensor mount drift after maintenanceExtrinsic yaw/pitch offsetCalibration residual detects drift; vehicle held for maintenance
Compute overload during event burstPerception node delayed while logging full-fidelity eventWatchdog/degraded mode engages before stale outputs are consumed

Evidence Artifacts

ArtifactContents
Fault matrixFault type, severity, seed, dataset, expected response
Corruption implementation reportWhere the fault is injected, validation of physical plausibility, known limitations
Run manifestBuild/map/calibration/log IDs and runtime parameters
Metric reportClean vs corrupted metrics, confidence intervals, failed slices
Safety action reportMonitor alerts, degraded-mode commands, stops, quarantines, operator notifications
Failure packetMinimal replay data, screenshots/plots, root cause, defect ID
Release dispositionPass/block/inconclusive decision and residual risk

Owner Handoffs

OwnerResponsibility
V&V robustness leadCampaign design, fault matrix, pass/block decision
Perception/SLAM ownerFault hooks, metric implementation, failure triage
Runtime assurance ownerMonitor response, degraded-mode verification, watchdog tests
Mapping ownerStale/wrong/shifted map injections and quarantine workflow
Data platform ownerReplay data, manifests, logs, evidence retention
Fleet operationsShadow-mode execution and operational safety controls
Safety boardResidual risk acceptance and ODD restrictions

Sources

Public research notes collected from public sources.