Skip to content

Key Decision Framework

Technology Choices and Their Rationale — Derived from 167 Research Documents


How to Use This Document

For each major architectural decision, this document provides:

  • The question being decided
  • Options considered (with evidence from the corpus)
  • Recommendation with rationale
  • Where to read more (corpus cross-references)

Decision 1: World Model Architecture

Question: Which world model approach to use for predicting future airside scenes?

OptionEvidenceVerdict
OccWorld (VQ-VAE + GPT transformer)Open-source, ECCV 2024, proven on nuScenes. mmdet3d dependency is painful but workable.Phase 1
Drive-OccWorld (action-conditioned)+33% over UniAD, action conditioning enables planning. AAAI 2025.Phase 2 upgrade
FlashOcc + custom forecasting197.6 FPS — only method viable on Orin real-time without optimization. But no forecasting built-in.Alternative Phase 1
Cosmos (latent diffusion)Commercially licensed, FSQ tokenizer, 14B params. Too large for on-vehicle.For simulation/data gen
Alpamayo VLA10B, camera-only, non-commercial license. Teacher model only.Phase 3 (distill)

Decision: Start with OccWorld for self-supervised LiDAR prediction (Phase 1). Upgrade to Drive-OccWorld when action-conditioned prediction is needed for planning (Phase 2). Use Cosmos for synthetic data generation (not on-vehicle). Skip VLA until Phase 3 when cameras + Thor hardware are available.

Read more: 30-autonomy-stack/world-models/occupancy-networks-comparison.md, 30-autonomy-stack/world-models/occworld-implementation.md, 30-autonomy-stack/vla-vlm/vla-distillation-scaling.md


Decision 2: Primary Perception Backbone

Question: What LiDAR 3D detection model to use?

OptionLatency (Orin)mAPLiDAR-only?Code
PointPillars6.84ms48%YesOpenPCDet
CenterPoint-Pillar~20ms50%YesOpenPCDet
CenterPoint-Voxel~40ms56%YesOpenPCDet
BEVFusion~40ms (25 FPS)68.5%No (camera+LiDAR)MIT BEVFusion

Decision: PointPillars for BEV feature extraction (6.84ms, well within budget). CenterPoint-Pillar for detection when object types matter (20ms). BEVFusion when cameras are added (Phase 2).

Key finding: PointPillars INT8 PTQ loses only 0.80% mAP for 2.2x speedup. QAT recovers to 0.17% loss.

Read more: 10-knowledge-base/geometry-3d/pointpillars.md, 30-autonomy-stack/perception/overview/openpcdet-centerpoint.md, 20-av-platform/compute/tensorrt-deployment-guide.md


Decision 3: Sensor Strategy

Question: Which sensors and in what order?

PhaseSensorsRationale
Phase 1Existing LiDAR (4-8 RoboSense)Already installed, proven, no hardware changes
Phase 1++ ADS-B receiver ($30)Jet blast hazard mapping, aircraft awareness
Phase 2+ Cameras (6-8 surround)Open-vocab detection, semantic understanding, VLA path
Phase 2+ 4D radar (2-4 Continental ARS548)PRIMARY adverse weather sensor, Doppler velocity
Phase 2++ Thermal/LWIR camera (FLIR Tura)Ground crew detection at night (hi-vis paradox mitigation)
Phase 3+ UWB beacons (near terminals)GPS-degraded area localization

Key finding: 4D radar should be PRIMARY, not backup — it's the only sensor immune to all airside adverse conditions (rain, fog, de-icing, jet exhaust). Cost: $50-200 per unit.

Read more: 20-av-platform/sensors/4d-radar.md, 60-safety-validation/verification-validation/robustness/airside-adverse-conditions.md, 20-av-platform/sensors/robosense-lidar.md


Decision 4: Compute Platform

Question: What hardware to run the new stack on?

PlatformTOPSCostAvailabilityCan Run
Jetson AGX Orin 64GB275 (sparse)$1,999NowPointPillars + OccWorld Lite + safety monitor
Jetson AGX Orin Industrial248HigherNowSame, -40C to +85C, 10yr lifecycle
DRIVE AGX Thor~1,000 (dense)TBD2025+Full world model + distilled VLA
Dual Orin550$4,000NowOne per stack (Simplex)

Decision: Orin 64GB for Phase 1-2 (sufficient for PointPillars + 50-200M world model + safety monitor). Thor for Phase 3 when VLA distillation is needed. Consider dual Orin if running both stacks requires GPU isolation.

Key finding: DLA contributes 74% of Orin compute at 15W, 3-5x more power-efficient than GPU. Use DLA for PointPillars, GPU for world model.

Read more: 20-av-platform/compute/nvidia-orin-technical.md, 20-av-platform/compute/nvidia-drive-thor.md, 20-av-platform/compute/tensorrt-deployment-guide.md


Decision 5: Map Strategy

Question: HD maps (current approach) or map-free (world model approach)?

ApproachDeploy TimeCost/AirportAccuracyMaintenance
HD map (current)3-6 months$50-200K surveycm-levelManual updates
AIXM + online perception1-2 weeks~$0 (AIXM free)m-level global, cm localSelf-updating
Pure map-freeDays$0Perception-limitedN/A

Decision: Keep HD maps for Phase 1 (proven, what reference airside AV stack does). Develop AIXM + online perception in parallel (Phase 2). Transition when world model accuracy validated. Keep HD map as fallback.

Read more: 30-autonomy-stack/localization-mapping/maps/map-free-driving.md, 10-knowledge-base/robotics/lanelet2-maps.md, 30-autonomy-stack/localization-mapping/overview/mapping-and-localization.md


Decision 6: Safety Architecture

Question: How to ensure the new stack doesn't compromise safety?

Decision: Simplex architecture — no question. Current stack becomes verified fallback. New stack is high-performance controller. Safety monitor arbitrates.

ComponentApproachRationale
Fallback controllerExisting reference airside AV stack (unchanged)Already production-tested
Safety monitorEnsemble OOD + RSS envelope + occupancy collisionMulti-layer, any failure → fallback
RSS parametersAirside-specific (2m aircraft, 3m personnel, 1s response)Conservative for airport
ArbitrationState machine with 2s hysteresisPrevents rapid switching
Hardware safetyIndependent safety PLC/MCU (comma.ai panda pattern)ML cannot override hardware safety

Key finding: comma.ai's panda safety layer (STM32H725, MISRA C, 100% line coverage + mutation testing) is the gold standard for hardware-independent safety. Safety must NEVER depend on neural networks.

Read more: 60-safety-validation/runtime-assurance/simplex-safety-architecture.md, 60-safety-validation/standards-certification/iso-3691-4-deep-dive.md, 80-industry-intel/companies/comma-ai/openpilot-codebase-analysis.md


Decision 7: Deployment Strategy

Question: Teleoperation-first (Fernride) or autonomy-first (TractEasy)?

ApproachRiskSpeed to RevenueTechnology Required
Autonomy-firstHigh (certification 12-24 months)SlowFull perception + planning + safety case
Teleoperation-firstLow (operator always in loop)FastConnectivity + remote UI + basic autonomy
Shadow mode firstNone (existing stack drives)No revenueSensors + compute + data pipeline

Decision: Shadow mode first (zero regulatory risk, builds data). Then teleoperation for revenue (Fernride model, operator ratio 1:4→1:10). Then graduated autonomy as world model matures.

Read more: 60-safety-validation/verification-validation/shadow-mode.md, 80-industry-intel/companies/fernride/tech-stack.md, 80-industry-intel/regulations/regulatory-trajectory-deep-dive.md


Decision 8: Certification Path

Question: What standard to certify against?

Decision: ISO 3691-4:2020 (what TractEasy and reference airside AV stack already use). Supplement with ISO/PAS 8800 (new AI safety lifecycle) and SOTIF (ISO 21448) for world model components.

StandardWhat It CoversCostTimeline
ISO 3691-4Driverless industrial trucks (primary)$130-380K12-24 months
ISO/PAS 8800AI safety lifecycle (supplement)Included aboveSame
ISO 21448 (SOTIF)Safety of intended functionalityIncludedSame
UL 4600Safety case framework$50-100K6-12 months

Key finding: ISO 3691-4 was harmonized with EU Machinery Directive in May 2024. New EU Machinery Regulation 2023/1230 (effective January 2027) mandates third-party assessment for AI autonomous vehicles.

Read more: 60-safety-validation/standards-certification/iso-3691-4-deep-dive.md, 60-safety-validation/standards-certification/certification-guide.md, 80-industry-intel/regulations/regulatory-trajectory-deep-dive.md


Decision Summary

#DecisionChoicePhase
1World modelOccWorld → Drive-OccWorld → VLA distill1 → 2 → 3
2PerceptionPointPillars (6.84ms) → BEVFusion1 → 2
3SensorsLiDAR → + 4D radar + cameras → + thermal + UWB1 → 2 → 3
4ComputeOrin 64GB → Thor1-2 → 3
5MapsHD map → AIXM + online perception1 → 2+
6SafetySimplex + hardware safety PLCAll phases
7DeploymentShadow → teleop → graduated autonomySequential
8CertificationISO 3691-4 + ISO/PAS 8800 + SOTIFStart now

Each decision is backed by evidence from the research corpus. See cross-references for full analysis.

Public research notes collected from public sources.