Research Methodology

How This Corpus Was Created

---

Overview

This research corpus began as a 24-hour intensive research session on 2026-03-21/22 using Claude Opus 4.6 with a 1M context window, augmented by parallel web-searching research agents. It has since been expanded and reorganized into an 866-document research corpus, surfaced as 870 VitePress reader pages.

The current reading surface is the static VitePress portal at https://kvynlim.github.io/industry-research/. The Markdown files remain the source of truth; the site adds search, generated navigation, clean URLs, last-updated metadata, and browser-friendly reading across 390k+ reader Markdown lines.

Domain stance: the corpus is a generic AV knowledge base. Airside autonomous vehicles are the most developed reference ODD because the current research base is deepest there, but method ratings and generic stack pages should not treat airside as the default deployment context.

Research Process

Phase 1: Foundational Research (20 parallel agents)

• Method: 20 specialized research agents launched simultaneously, each covering a distinct topic domain
• Scope: World models, VLAs, diffusion models, RL, 3DGS/NeRF, occupancy prediction, JEPA, LLM reasoning, motion prediction, multi-agent coordination, datasets/benchmarks, safety certification, compute hardware, perception foundation models, mapping/localization, robustness, and company strategies
• Sources: WebSearch across academic papers (arXiv, conference proceedings), company websites, developer documentation, GitHub repositories, press releases, regulatory documents
• Output: 20 research reports, ~13,000 lines

Phase 2: Design & Brainstorming

• Method: Explored reference AV stack patterns and operational constraints to understand the target system architecture
• Output: Design specification (891 lines, spec-reviewed with automated review agent), POC proposals, master synthesis

Phase 3: Execution Guides (20 parallel agents)

• Method: 20 agents focused on practical, implementation-ready knowledge
• Topics: BEV encoding, OccWorld setup, data engine from ROS bags, Simplex safety architecture, MLOps, Alpamayo, Cosmos, 3DGS digital twin, airport data APIs, transfer learning, ROS 2 migration, Frenet planner augmentation, OpenPCDet/CenterPoint, Dreamer RL, open-source ecosystem, shadow mode, FOD/jet blast, turnaround prediction, open-vocab detection, E2E pipeline
• Output: 20 execution guides, ~13,000 lines

Phase 4: Production Deployment (15 parallel agents)

• Method: 15 agents researching real-world deployment case studies
• Companies: TractEasy, comma.ai, Waymo, Tesla, Changi programme
• Topics: Safety incidents, OTA fleet management, safety certification, teleoperation, 5G connectivity, production perception, Moonware HALO, deployment playbook, production ML deployment
• Output: 15 deployment reports, ~13,000 lines

Phase 5: First Principles (6 written + 9 agents)

• Method: Deep first-principles derivations for core technologies
• Topics: PointPillars (tensor shapes), VQ-VAE/FSQ tokenization, transformer world models, bicycle kinematic model, diffusion models, RTK-GPS/IMU localization, GTSAM factor graphs, Lanelet2 maps, Frenet trajectory math, CAN bus DBW, Mamba SSM
• Output: 11 foundation documents, ~7,000 lines

Phase 6: Gap-Filling Deep Dives (20 parallel agents)

• Method: Targeted deep dives on specific gaps identified in the corpus
• Topics: ISO 3691-4 certification, Waymo safety methodology, airport data systems (real API endpoints), nuScenes/Waymo practical guide, TensorRT deployment, Autoware Universe, airport 5G case studies, Mamba SSM, ground crew safety, occupancy networks comparison (20 methods), simulators for airside, regulatory trajectory, open-source world model repos (21 evaluated), pushback systems, insurance/liability, comma.ai codebase analysis, 4D radar, DINOv2 for driving, airport digital twins, fleet management dispatch
• Output: 20 deep-dive reports

Phase 7: Restructuring & Synthesis

• Method: Reorganized from the early flat/topic-based corpus into the final numbered end-to-end knowledge architecture: 00-start-here/, 10-knowledge-base/, 20-av-platform/, 30-autonomy-stack/, 40-runtime-systems/, 50-cloud-fleet/, 60-safety-validation/, 70-operations-domains/, 80-industry-intel/, and 90-synthesis/.
• Output: Root navigation, competitive landscape, technology readiness, getting-started guide, risk register, cross-references, and numbered reader paths

Phase 8: Method-Level SLAM Expansion and Coverage Audit

• Method: Parallel web-search agents audited LiDAR, visual, dense/RGB-D, LiDAR-visual-inertial, radar, registration, loop-closure, and backend SLAM coverage against the existing method library
• Output: Dedicated GLIM method file plus SLAM Coverage Audit and Backlog, with P0/P1/P2 missing-method queues, 2026-05-08 latest-method and gap-discovery sweeps, and source links

Phase 9: Perception Stack Coverage Audit

• Method: Multiple rounds of parallel research agents audited camera BEV, occupancy, LiDAR/radar/thermal/event perception, open-world/OOD perception, temporal tracking, cooperative/V2X perception, robustness, deployment validation, and benchmarks
• Output: Perception Coverage Audit and Backlog, with P0/P1/P2 missing-method queues, benchmark gaps, discoverability fixes, and source links

Phase 10: Method-Level Perception Library

• Method: Five parallel writing agents split the perception coverage audit into atomic, one-method research files across camera BEV/occupancy, LiDAR/radar/event/FMCW perception, open-world/open-vocabulary perception, robust fusion/validation, and cooperative/latency/data-engine methods
• Output: Perception Method Library, initially with 54 single-technique method files and now expanded to 138 method-library files, including 137 atomic method pages plus the overview, that follow a shared structure for core idea, inputs/outputs, architecture, training/evaluation, strengths, failure modes, domain fit, transfer notes for explicitly scoped ODDs, implementation notes, and sources

Phase 11: Cross-Architecture Knowledge Gap Audit

• Method: Six parallel research agents audited the post-restructure architecture across foundations, AV platform, autonomy stack, runtime/cloud, safety/validation, and operations/industry. One autonomy agent was split into two narrower replacement agents after exceeding context, preserving coverage without overloading the review.
• Output: End-to-End AV Knowledge Gap Backlog, with P0/P1/P2 missing-file queues across reusable fundamentals, platform power/thermal/diagnostics, planning/control/V2X, runtime/cloud operations, safety evidence, and non-airside operations domains.

Phase 12: P0 Knowledge Gap Research Wave

• Method: Six parallel writing agents plus one follow-up agent converted the P0 gap backlog into first-class research files. Each agent owned a disjoint write scope: foundations, AV platform, planning/control/V2X, E2E/VLA/world models, runtime/cloud+safety evidence, operations domains, and delivery robots.
• Output: 35 source-backed P0 gap files across 10-knowledge-base/, 20-av-platform/, 30-autonomy-stack/, 40-runtime-systems/, 50-cloud-fleet/, 60-safety-validation/, and 70-operations-domains/.

Phase 13: Perception, SLAM, and Sensor Deep-Dive Loop

• Method: Parallel discovery agents re-audited perception, SLAM, Gaussian/3DGS methods, and sensor fundamentals, then six writing agents promoted the highest-value gaps into atomic files. The wave explicitly covered SplatAD and other Gaussian/4DGS methods, production-relevant LIVO/SLAM stacks, radar/Gaussian SLAM, and sensor measurement/noise models for perception, SLAM, and mapping.
• Output: 33 source-backed files: 9 perception method pages, 13 SLAM method pages, 9 knowledge-base sensor/state-estimation fundamentals, 2 platform sensor hardware pages, plus a Continuous Research Loop to keep the next gap queue active.

Phase 14: First-Principles Foundations Expansion

• Method: Five parallel web/discovery rounds audited probability/statistics, nonlinear optimization, numerical linear algebra, association/tracking, and broader AV robotics foundations. Five writing agents then promoted the selected gaps into atomic first-principles KB files with disjoint ownership.
• Output: 33 source-backed knowledge-base files across probability/statistics, optimization, numerical linear algebra, geometry, mapping, state estimation, sensors, signal processing, and systems engineering. The wave covers Gaussian noise, Mahalanobis gating, MAP/MLE, robust statistics, mixtures, Gauss-Newton, Levenberg-Marquardt, Cholesky, QR/SVD, sparse solvers, Lie groups, PnP, ICP/GICP/NDT, occupancy grids, data association, JPDA/MHT/RFS, filters, sensor likelihoods, radar ambiguity, CFAR, timestamping, and statistical benchmarking.

Phase 15: LIORNet, LiDAR Removal, and Machine-Learning Foundations

• Method: Parallel discovery agents audited LIORNet, adverse-weather LiDAR denoising, classical outlier removal, map-cleaning methods, weather datasets, and first-principles ML gaps. Five writing agents then promoted the selected work into disjoint file groups: learned LiDAR denoisers, broad removal and map-cleaning techniques, weather robustness datasets, classical ML foundations, and modern ML foundations.
• Output: 41 source-backed files: LIORNet and adjacent denoising methods, classical LiDAR outlier and weather artifact removal, LiDAR ghost/multipath artifacts, artifact-removal validation, ERASOR/Removert/map-cleaning pages, weather robustness dataset pages, and a machine-learning ladder from perceptrons, logits, backprop, optimization, CNNs, and RNNs to transformers, Mamba, JEPA, foundation-model training, and world-model first principles.

Phase 16: Dynamic/Static Object Removal and ML Objective Foundations

• Method: Parallel discovery and writing agents expanded the removal topic from weather/noise filtering into dynamic-object removal, static-but-wrong-object removal, scene flow, MOS, map-change datasets, and map-cleaning benchmarks. A parallel ML wave filled first-principles gaps around representation objectives, EBMs, masked modeling, diffusion/flow sampling, tokenization, positional encodings, calibration, leakage, multi-task losses, and world-model evaluation.
• Output: 26 source-backed files: MapCleaner, ERASOR++, 4dNDF, FreeDOM, STATIC-LIO dynamic-point removal, MotionSeg3D, MambaMOS, neural scene-flow priors, moving/static separation datasets, moved-object map-change datasets, 4D occupancy and scene-flow benchmarks, an airside dynamic map-cleaning benchmark, and 11 machine-learning foundation notes that bridge classical neural-network training to modern transformer, Mamba, diffusion, JEPA, and world-model pipelines.

Phase 17: Perception, SLAM, KB, and Validation Web-Gap Expansion

• Method: Five web-search scout agents re-audited perception, SLAM, world-model/neural-field, dataset/validation, and knowledge-base gaps, then six writing agents promoted the highest-value gaps with disjoint ownership.
• Output: 31 source-backed files: 9 perception method pages, 2 world-model pages, 9 SLAM/localization pages, 5 knowledge-base probability/control foundation notes, 4 perception dataset/benchmark pages, and 2 validation protocol pages. The wave added CVFusion, 4D radar-camera occupancy, FMCW LiDAR predictive detection, cross-domain scene flow, TrackOcc, DrivingGaussian, HUGS, SplatFlow, DistillNeRF, self-supervised occupancy flow, UniScene, robust/certifiable PGO, Kimera-RPGO/PCM, distributed multi-robot PGO, LT-mapper/Khronos, RTMap/DUFOMap, GPR localization, radar teach-repeat, MOVES, probabilistic graphical models, information theory, calibration/conformal uncertainty, constrained MPC/iLQR, MDP/POMDP foundations, MUSES, corruption/OOD/FOD benchmarks, FOD validation, and knowledge-base evaluation.

Phase 18: Perception, SLAM, First-Principles, and Reliability Gap Loop

• Method: Six web-search scouts compared the repo against 2024-2026 perception, SLAM, knowledge-base, dataset, and validation sources. Six writing agents then promoted disjoint file sets directly into the main tree.
• Output: 36 source-backed files: 12 perception method/dataset pages, 12 SLAM/localization pages, 6 first-principles knowledge-base pages, 5 safety-validation protocols, and 1 fleet-data contract. The wave added LiDAR-camera occupancy fusion, dynamic occupancy/free-space, radar-LiDAR adverse-weather detection, RobuRCDet, SAMFusion, spatiotemporal occupancy flow, STU 3D anomaly segmentation, synthetic multimodal FOD benchmarks, OVAD/OVODA, open-vocabulary panoptic occupancy, RCP-Bench, V2X large-range sequential datasets, Scan Context, LiDAR bundle-adjustment factors, Kimera-Multi, COVINS/COVINS-G, D2SLAM, UWB/range SLAM, OKVIS2-X, MM-LINS, event-camera VIO/SLAM, thermal-inertial SLAM, 4D imaging-radar RIO, radar-to-LiDAR map localization, continuous-time trajectory priors, motion distortion, volumetric map representations, volume rendering/3DGS foundations, detection operating points, tracking lifecycle metrics, perception/SLAM/map evidence cases, statistical-validity protocols, uncertainty calibration release gates, corruption fault injection, SLAM/map benchmark protocols, and the perception/SLAM fleet-data contract.

Phase 19: Recurring Research-Maintenance Loop

• Method: One bounded batch per wake, with five read-only research scouts per pass and local integration after source verification, deduplication, cross-linking, and repo checks.
• Latest loop addition: The recurring loop is now 40+ source-backed files plus source-backed refreshes; the newest batches refreshed the aggregated-map segmentation dataset/proxy landscape, routed Point Cloud City / Open3D-ML PCC and City-Facade as managed-building/public-safety and MLS facade/vertical-structure proxies, routed ZAHA as a WACV 2025 / TUM2TWIN facade-hierarchy benchmark with 601M annotated points and LoFG2/LoFG3 labels for terminal-frontage and digital-twin taxonomy stress, promoted GridNet-HD Power-Line LiDAR-Image Segmentation as a utility-infrastructure LiDAR-image benchmark for long-thin classes, added a modality-aware backbone/deployment-contract comparison for end-to-end aggregated-map semantic segmentation, added checked JSON Schema contracts and examples for semantic-map manifests plus runtime map contracts, extended the manifest prior-input contract for semantic-SLAM histograms, neural/Gaussian map priors, and static/dynamic masks, promoted static-map lifecycle coverage through Lifelong 3D Map Version Control, Potentially Dynamic Object Removal by Ground Projection, Uni-Mapper Dynamic-Aware LiDAR Map Merging, LAMM Multi-Session Point-Cloud Map Merging, BeautyMap, and Raymoval, expanded ML-related SLAM coverage through Semantic SLAM, Dynamic-Object-Aware SLAM, Multi-Agent Neural Gaussian SLAM, and Object-Level SLAM, added Point-Cloud Mamba / SSM Backbones as a research-stage efficiency branch for aggregated-map segmentation, promoted LOSC as an atomic open-vocabulary LiDAR pseudo-label consolidation method, refreshed the map-hygiene ground-truth protocol workflow across semantic-map QA, publication gates, monitoring, and safety evidence, added a compact proxy/input/training selector for end-to-end aggregated-map semantic segmentation, and added an explicit FOD/adverse-airside proxy-boundary refresh across existing dataset, validation, and removal routes.
• Latest map-conditioning refresh: OpenLiDARMap and the complementary FlexCloud branch were routed through the HD map-construction pipeline, aggregated-map source conditioning, source-registry, README, INDEX, and methodology surfaces as georeferenced point-cloud map-conditioning references for GCP-sparse or GNSS-assisted map construction, not as semantic segmentation datasets or label sources. The following bounded pass promoted LAMM Multi-Session Point-Cloud Map Merging as the large-scale multi-session point-cloud map-merging branch that conditions the geometric substrate before semantic segmentation.
• Latest map-quality refresh: MapEval Point-Cloud Map-Quality Evaluation was promoted as the direct source-map geometry QA gate after map merging/dynamic cleanup and before aggregated-map semantic segmentation, localization regression, or map publication. LEMON-Mapping remains a watchlist/comparison item until official code, license, or publication maturity improves.
• Output: 40+ source-backed files plus queue cleanups and source-backed refreshes: DepthOcc, COSMO-Bench, SparseBEV, DETR4D, TEOcc, GaussRender, DSERT-RoLL, CMHT, SLAM Toolbox, NDT variants/NDT maps, LaMAria, Hilti x Trimble SLAM Challenge 2026, ScaleMaster, HI-SLAM2, SEGS-SLAM, Neural/Gaussian SLAM Surveys, GV-iRIOM, GVINS/GLIO raw GNSS factor fusion, Wheel Odometry and Vehicle Motion Factors, CAO-RONet, RKO-LIO, Doppler Radar-LiDAR SLAM, Radar RIO Correspondence and Uncertainty, Lifelong 3D Map Version Control, Potentially Dynamic Object Removal by Ground Projection, Uni-Mapper Dynamic-Aware LiDAR Map Merging, LAMM Multi-Session Point-Cloud Map Merging, MapEval Point-Cloud Map-Quality Evaluation, BeautyMap, Raymoval, Point-Cloud Mamba / SSM Backbones, LOSC, GaussianFlowOcc, GaussTR, GS-Occ3D, Calibration Bay Fixtures, QuantV2X, SparseCoop, TruckV2X, VOGS-CP, Airport-FOD3S Synthetic FOD Data Engine, EmbodiedScan/MMScan Embodied 3D Benchmarks, SpaCeFormer, Active Calibration Experiment Design, Robust-Loss Covariance Consistency, Epipolar Geometry/Homographies Two-View Verification, Optical Flow and Scene Flow First Principles, Constrained KKT/QP/SQP Solver Mechanics, AV Data Evaluation Fundamentals, Ultrasonic Proximity Sensing Models, Thermal IR Radiometry First Principles, Infrastructure-Aided Localization, and Fiducial and Corner Localization. The loop also refined aggregated-map semantic segmentation dataset/proxy routing, 4D radar freespace and dynamic occupancy coverage, corrected SLAM Toolbox licensing/discoverability, broadened radar-inertial online calibration to include spatio-temporal calibration, refreshed Drive-OccWorld/world-model follow-on routing, corrected LinkOcc source routing, routed Oxford Spires/IILABS 3D/SMapper-light/FusionPortableV2/S3E into the SLAM benchmark matrix with caveats, corrected count drift and duplicate-prone queue wording, tightened world-model metric claims, promoted GS-Occ3D as vision-only Gaussian-surfel occupancy label curation, added cooperative-perception quantization coverage through QuantV2X, added a sparse cooperative-query baseline through SparseCoop, added TruckV2X as a truck-centered V2X dataset proxy for articulated vehicles and long GSE, promoted VOGS-CP as a collaborative Gaussian semantic-occupancy method, promoted Airport-FOD3S as a synthetic FOD data-engine workflow with DualFOD and LIDAROC routing, promoted GVINS/GLIO as a raw pseudorange/Doppler GNSS factor-fusion page, promoted wheel/vehicle-motion factors as a slip-aware SLAM/localization factor-family page, promoted CAO-RONet as a learned 4D radar odometry baseline, promoted RKO-LIO as a robust sensor-agnostic LIO baseline, promoted Doppler Radar-LiDAR SLAM as a bridge page for Radarize, DRO, and Doppler-SLAM, promoted Radar RIO Correspondence and Uncertainty as a combined radar-inertial association/confidence page, refreshed radar place-recognition descriptor routing with TransLoc4D, 4D RadarPR, TDFANet, and DIDLM dataset caveats, promoted BeautyMap as a training-free dynamic-object-removal page for binary visibility maps, KTH DynamicMap benchmarking, and aggregated-map semantic-label cleanup, promoted Raymoval as a paper-backed azimuth-elevation raycasting cleaner with ERASOR-lineage metric caveats, promoted LAMM as the multi-session point-cloud map-merging branch after Uni-Mapper, promoted MapEval as the direct point-cloud map-quality QA branch before semantic labels are trusted, promoted EmbodiedScan/MMScan as an embodied 3D spatial-grounding benchmark reference, refreshed SparseDriveV2/DiffusionDriveV2 planning and E2E benchmark routing with PDMS/EPDMS separation, routed DySS through sparse-query and streaming-temporal overviews pending official code/adoption, promoted SpaCeFormer as a fast proposal-free open-vocabulary 3D instance segmentation reference, promoted point-cloud Mamba/SSM backbones as a map-scale token-cost research branch rather than a production default, promoted LOSC as the open-vocabulary LiDAR pseudo-label consolidation page, refreshed map-hygiene ground-truth protocol ownership across semantic-map QA, map construction, publication gates, monitoring, and regulatory evidence, promoted active calibration experiment design as the first-principles bridge from observability metrics to safe data collection, promoted robust-loss covariance consistency as the bridge from robust weighting/gating to covariance reporting and release evidence, promoted epipolar/homography two-view verification as the bridge from image correspondences to visual odometry, stereo, loop closure, and calibration QA, promoted optical/scene-flow foundations as the bridge from pixel and point motion fields to dynamic occupancy, map cleaning, and flow-aware release evidence, promoted constrained solver mechanics as the bridge from KKT/QP/SQP theory to MPC, CBF-QP, and constrained trajectory-release evidence, promoted AV data evaluation fundamentals as the split, coverage, benchmark-ladder, and release-artifact contract for learned autonomy evidence, refreshed sensor calibration fleet operations around package lifecycle, manifests, telemetry schemas, quarantine, rollback, and SUMS-aligned release evidence, and recorded that no public dataset found in this pass directly covers airside dust, de-icing mist, steam, glycol film, or wet-apron multipath.
• Most recent maintenance refresh: The 2026-05-24 loop promoted MapEval as a source-backed point-cloud map-quality evaluation page, routed it through SLAM benchmarking, map construction, aggregated-map semantic QA, map-cleaning, LAMM/Uni-Mapper, source-registry, README, INDEX, and methodology surfaces, then tightened the follow-on QA contract so metrics_evidence.qa_report_id dereferences to a source_map_quality evidence block and the runtime-map example follows Autoware's 20 m divided-point-cloud convention. The same continuing loop expanded the MLOps scale track with evaluation/replay gates, registry lifecycle, serving/inference operations, platform SRE/reliability, and data catalog/lineage/quality operations by scale, cross-linking serving manifests, endpoint/batch/edge rollout, platform SLOs, error budgets, backup/restore, DR, tenant isolation, data-product contracts, OpenLineage-style event boundaries, quality gate severity, deletion propagation, ODD-cell canary, rollback, and incident lanes through runtime deployment, registry, evaluation, monitoring, governance, attestation, OTA compatibility, README, INDEX, and methodology. LEMON-Mapping remains a watchlist/comparison item pending stronger source/code maturity; II-NVM and DQFormer are queued as source-mature future candidates.

Quality Controls

1. Spec Review: Design specification reviewed by automated spec-review agent with factual corrections
2. Factual Corrections: Known inaccuracies identified and propagated across all documents (V-JEPA 240x→~15x, Alpamayo naming/licensing, Copilot4D parameter count)
3. Cross-Referencing: Synthesis documents cross-reference each other and the detailed research
4. Source Attribution: Each research document includes a Sources section with paper references, URLs, and datasets
5. Primary-Source Preference: Implementation claims are checked against primary artifacts where available, with company-specific claims kept attributed
6. Coverage Audits: Broad method libraries now include explicit backlog documents for missing first-class pages, starting with SLAM and perception
7. Atomic Method Pages: SLAM and perception now separate overview synthesis from one-method research files, so individual techniques can be updated and compared without burying them inside family documents
8. Cross-Architecture Gap Tracking: The synthesis layer now tracks P0/P1/P2 research gaps outside the dedicated SLAM and perception audits
9. P0 Gap Promotion: High-priority cross-architecture gaps are promoted into first-class files before P1/P2 backlog work begins
10. Continuous Research Loop: Discovery, triage, promotion, cross-linking, verification, and next-queue selection are now documented as a repeatable loop
11. First-Principles Layering: Applied perception, SLAM, mapping, and sensor files now link back to reusable math primitives instead of repeating estimator fundamentals inline
12. Removal Safety Separation: LiDAR artifact removal now separates nuisance-point deletion, ghost/multipath diagnosis, dynamic-map cleaning, and safety validation so filtering does not become an unexamined hazard-deletion step
13. Domain Fit Rebalance: Generic method and synthesis pages should use Domain Fit language across canonical AV domains, while airside-specific pages remain airside-first with transfer notes where relevant.

Limitations

1. Web search rate limits: Some agents hit API rate limits during research. Affected topics were written from training knowledge rather than live web search.
2. Point-in-time: Research broadly reflects the state of the field as of March 2026, with 2026-05-08, 2026-05-09, and 2026-05-22/23/24 refreshes for SLAM, perception, Gaussian/3DGS methods including VOGS-CP, sensor fundamentals, first-principles estimator math including active calibration experiment design, robust-loss covariance consistency, epipolar/homography two-view verification, optical/scene-flow motion fields, and constrained KKT/QP/SQP solver mechanics, LIORNet/adverse-weather LiDAR removal, dynamic/static object removal including BeautyMap and Raymoval, scene-flow/MOS benchmarks, moved-object map-change datasets, weather datasets including DSERT-RoLL, CMHT, and LIDAROC proxy routing, machine-learning foundations including AV data evaluation fundamentals, radar-camera/FMCW perception, robust SLAM backends, collaborative and alternative-sensor SLAM, raw GNSS factor fusion through GVINS/GLIO, wheel/vehicle-motion factors, sensor-agnostic LiDAR-inertial odometry through RKO-LIO, learned 4D radar odometry through CAO-RONet, radar RIO correspondence/uncertainty, Doppler radar-LiDAR SLAM through Radarize/DRO/Doppler-SLAM routing, neural planning and E2E benchmark refreshes through SparseDriveV2/DiffusionDriveV2, lifelong localization, adverse/OOD/FOD/V2X benchmarks including TruckV2X and DualFOD/FOD-UAS, Airport-FOD3S synthetic data-engine coverage, embodied 3D benchmarks including EmbodiedScan/MMScan, open-vocabulary 3D instance segmentation through SpaCeFormer, and perception/SLAM reliability protocols. Fast-moving areas (world models, VLAs, neural/Gaussian SLAM, neural planning, open-world perception, dynamic map cleaning, adverse-weather denoising, embodied 3D grounding, and 4D radar) may have newer developments. The 2026-05-23 recurring passes also added ultrasonic proximity sensing models and thermal IR radiometry to the sensor-fundamentals refresh, refreshed fleet calibration operations with package lifecycle, telemetry-schema, quarantine, rollback, and release-evidence controls, promoted infrastructure-aided localization for managed-site UWB, fiducial, RFID/BLE, Wi-Fi RTT, magnetic-map, reflector, and 5G NR/mmWave aids, and split fiducial/corner localization into a geometry foundation page.
3. Reference-ODD imbalance: Airside has the deepest current coverage and no public large-scale airside driving datasets, so some deployment comparisons rely on published reports rather than reproducible benchmarks. Generic method pages should separate airside-specific evidence from broader AV deployment relevance.
4. Company information: Some companies (UISEE, AeroVect) have limited public technical information. Claims are attributed but not all independently verified.
5. Regulatory predictions: Timeline predictions for FAA/EASA standards are based on published roadmaps and industry trends, not official commitments.

Corpus Statistics

Metric	Value
Core research documents	852
Reader pages	856
Total lines	390k+
Research agents spawned	300+
Companies researched	25
Method-level SLAM library	158 SLAM-method documents including overview/audit
Method-level perception files	138
Papers referenced	700+
GitHub repos evaluated	90+
API endpoints documented	15+
Airport deployments documented	15+
Static reader	VitePress on GitHub Pages
Initial research sprint	~24 hours

How to Extend This Research

1. Add a new company: Create 80-industry-intel/companies/<name>/tech-stack.md, update INDEX.md and README.md
2. Add new platform research: Create it in the appropriate 20-av-platform/<domain>/ directory for compute, sensors, networking/connectivity, drive-by-wire, power/electrical, diagnostics, ruggedization, or thermal material.
3. Add new autonomy-stack research: Create it in the appropriate 30-autonomy-stack/<domain>/ directory for world models, perception, planning, localization/mapping, simulation, VLA/VLM, E2E driving, or multi-agent/V2X material.
4. Add safety, validation, or robustness research: Create it in the appropriate 60-safety-validation/<domain>/ directory.
5. Add operational or industry research: Use 70-operations-domains/ for domain operations across airside, warehouse, logistics yard, port, mining, construction, agriculture, road AV, delivery robot, and outdoor campus material. Use 80-industry-intel/ for companies, market intelligence, regulations, and cross-domain deployment evidence.
6. Update a finding: Edit the document, run rg to find all references to the finding across the corpus, update all
7. Add a new POC: Add to 90-synthesis/poc-roadmaps/poc-proposals.md and 90-synthesis/readiness-risk/technology-readiness.md
8. Track regulatory changes: Update 80-industry-intel/regulations/regulatory-trajectory-deep-dive.md