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?
| Option | Evidence | Verdict |
|---|---|---|
| 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 forecasting | 197.6 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 VLA | 10B, 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?
| Option | Latency (Orin) | mAP | LiDAR-only? | Code |
|---|---|---|---|---|
| PointPillars | 6.84ms | 48% | Yes | OpenPCDet |
| CenterPoint-Pillar | ~20ms | 50% | Yes | OpenPCDet |
| CenterPoint-Voxel | ~40ms | 56% | Yes | OpenPCDet |
| 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?
| Phase | Sensors | Rationale |
|---|---|---|
| Phase 1 | Existing 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?
| Platform | TOPS | Cost | Availability | Can Run |
|---|---|---|---|---|
| Jetson AGX Orin 64GB | 275 (sparse) | $1,999 | Now | PointPillars + OccWorld Lite + safety monitor |
| Jetson AGX Orin Industrial | 248 | Higher | Now | Same, -40C to +85C, 10yr lifecycle |
| DRIVE AGX Thor | ~1,000 (dense) | TBD | 2025+ | Full world model + distilled VLA |
| Dual Orin | 550 | $4,000 | Now | One 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)?
| Approach | Deploy Time | Cost/Airport | Accuracy | Maintenance |
|---|---|---|---|---|
| HD map (current) | 3-6 months | $50-200K survey | cm-level | Manual updates |
| AIXM + online perception | 1-2 weeks | ~$0 (AIXM free) | m-level global, cm local | Self-updating |
| Pure map-free | Days | $0 | Perception-limited | N/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.
| Component | Approach | Rationale |
|---|---|---|
| Fallback controller | Existing reference airside AV stack (unchanged) | Already production-tested |
| Safety monitor | Ensemble OOD + RSS envelope + occupancy collision | Multi-layer, any failure → fallback |
| RSS parameters | Airside-specific (2m aircraft, 3m personnel, 1s response) | Conservative for airport |
| Arbitration | State machine with 2s hysteresis | Prevents rapid switching |
| Hardware safety | Independent 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)?
| Approach | Risk | Speed to Revenue | Technology Required |
|---|---|---|---|
| Autonomy-first | High (certification 12-24 months) | Slow | Full perception + planning + safety case |
| Teleoperation-first | Low (operator always in loop) | Fast | Connectivity + remote UI + basic autonomy |
| Shadow mode first | None (existing stack drives) | No revenue | Sensors + 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.
| Standard | What It Covers | Cost | Timeline |
|---|---|---|---|
| ISO 3691-4 | Driverless industrial trucks (primary) | $130-380K | 12-24 months |
| ISO/PAS 8800 | AI safety lifecycle (supplement) | Included above | Same |
| ISO 21448 (SOTIF) | Safety of intended functionality | Included | Same |
| UL 4600 | Safety case framework | $50-100K | 6-12 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
| # | Decision | Choice | Phase |
|---|---|---|---|
| 1 | World model | OccWorld → Drive-OccWorld → VLA distill | 1 → 2 → 3 |
| 2 | Perception | PointPillars (6.84ms) → BEVFusion | 1 → 2 |
| 3 | Sensors | LiDAR → + 4D radar + cameras → + thermal + UWB | 1 → 2 → 3 |
| 4 | Compute | Orin 64GB → Thor | 1-2 → 3 |
| 5 | Maps | HD map → AIXM + online perception | 1 → 2+ |
| 6 | Safety | Simplex + hardware safety PLC | All phases |
| 7 | Deployment | Shadow → teleop → graduated autonomy | Sequential |
| 8 | Certification | ISO 3691-4 + ISO/PAS 8800 + SOTIF | Start now |
Each decision is backed by evidence from the research corpus. See cross-references for full analysis.