Skip to content

Risk Register

Risks to POC Execution and World Model Integration


Risk Matrix

IDRiskLikelihoodImpactScoreMitigationOwner
R1Insufficient airside training dataMediumHighHighSelf-supervised training (no labels needed), transfer from nuScenes, synthetic data generationML Lead
R2World model doesn't generalize to airsideMediumHighHighPre-train on road data, progressive fine-tuning, extensive eval before deploymentML Lead
R3Orin compute insufficient for full pipelineLowHighMediumLite model tier (50M params), DLA offloading, cloud-edge hybrid for heavy modelsSystems Lead
R4FAA/airport authority blocks deploymentMediumHighHighStart with shadow mode (no regulatory barrier), build safety case early, engage FAA proactivelyProgram Lead
R5RoboSense LiDAR data incompatible with pretrained modelsLowMediumLowPoint cloud normalization pipeline, intensity/density matching, documented in data-engine guideML Lead
R6GPS multipath near terminals causes localization failureMediumMediumMediumLiDAR SLAM fallback already in GTSAM stack, UWB beacons as backupSystems Lead
R7Integration disrupts production reference airside AV stackLowCriticalHighSimplex architecture: new stack runs in parallel, never touches production until validatedSystems Lead
R8Key ML dependencies become unmaintainedMediumMediumMediumUse established frameworks (OpenPCDet, PyTorch), avoid single-source dependencies (OccWorld mmdet3d risk)ML Lead
R9Bag data is corrupted/incompleteLowMediumLowIndex all bags first (Day 1), validate quality before training, keep originalsData Lead
R10Camera hardware delaysMediumLowLowPhase 1-3 are LiDAR-only, cameras only needed for Phase 2+ (POC 7)Hardware Lead
R11Airport operations data (A-CDM) access deniedHighMediumHighStart with ADS-B (freely available), build relationship with airport ops teamProgram Lead
R12World model hallucinations cause unsafe behaviorLowCriticalMediumSimplex fallback, OOD detection, RSS safety envelope, safety monitor ensembleSafety Lead
R13Competitor (UISEE/TractEasy) achieves world model firstLowMediumLowNo competitor currently uses world models; UISEE is classical perceptionStrategy Lead
R14EU Machinery Regulation 2023/1230 (Jan 2027) imposes new requirementsHighMediumHighTrack regulation, ensure AI components have explainability and logging built inCompliance Lead
R15De-icing fluid/jet blast damages sensorsMediumMediumMedium4D radar as primary (weather-immune), sensor cleaning systems, thermal camerasHardware Lead
R16Insurance costs prohibitive for autonomous operationsMediumMediumMediumStart with shadow mode (standard insurance), Swiss Re data shows 88% fewer claims for AVProgram Lead
R17Team lacks ML engineering expertiseMediumHighHighStart with proven pipelines (OpenPCDet, OccWorld), use cloud GPU (Lambda Labs), hire/trainProgram Lead
R18Hi-vis clothing causes AEB failure at nightMediumCriticalHighAdd thermal/LWIR camera (FLIR Tura ASIL-B), UWB personal transponders, multi-layer detectionSafety Lead

Top 5 Risks (Action Required)

1. R7: Integration disrupts production stack (Impact: Critical)

Status: Mitigated by design Action: Simplex architecture ensures complete isolation. Shadow mode runs new stack without actuator access. Arbitrator node with hysteresis prevents rapid switching. Both stacks tested independently.

2. R12: World model hallucinations (Impact: Critical)

Status: Mitigated by design Action: Four-layer safety monitor (OOD detection, RSS envelope, occupancy collision check, watchdog). If ANY layer fails → fallback to production stack. All decisions logged for analysis.

3. R18: Hi-vis AEB failure at night (Impact: Critical)

Status: Requires hardware action Action: Add thermal/LWIR camera to sensor suite. FLIR Tura has ASIL-B rating. UWB personal transponders provide redundant crew detection. Seven-layer detection stack designed in ground-crew-pedestrian-safety.md.

4. R4: Regulatory blocking (Impact: High)

Status: Requires proactive engagement Action: Shadow mode requires NO regulatory approval (vehicle drives with current stack). Build ISO 3691-4 safety case in parallel ($130K-380K, 12-24 months). Engage FAA via CertAlert 24-02 dialogue channel. Track EASA AI Roadmap 2.0.

5. R17: ML expertise gap (Impact: High)

Status: Requires investment Action: POCs 4 and 5 (jet blast, FOD) need zero ML. POCs 1 and 2 use established frameworks with pretrained models. Getting-started guide provides Day 1 runnable code. Cloud GPU eliminates infrastructure barrier. Consider Aston University KTP extension to cover perception ML.


Risk Appetite Statement

Safety: Zero tolerance for unmitigated safety risks. Any system that could harm personnel, damage aircraft, or cause regulatory violation must have independent safety layers.

Technology: Moderate appetite for technology risk. Use proven open-source frameworks where available. Accept that world models are research-stage but mitigate with Simplex fallback.

Schedule: High appetite for schedule risk. Better to ship a working POC late than a broken one on time. Quality gates at each phase transition.

Cost: Low appetite for cost overruns. POC budget $2-5K (cloud GPU + ADS-B hardware). Certification budget $130-380K is separate and well-estimated.


Risk register should be reviewed monthly and updated as POCs progress. Risk scores recalculated after each phase gate.

Public research notes collected from public sources.