Skip to content

Industry Research

Markdown-first knowledge base for autonomous vehicle technology across road, airside, warehouse, logistics yard, port, mining, construction, agriculture, delivery robot, and outdoor campus domains. Airside autonomous vehicles remain the best-developed reference ODD, not the default evaluation lens.

Read it as a site: https://kvynlim.github.io/industry-research/

The repository remains Markdown-first, but the VitePress reader is the intended reading surface: local search, generated sidebar navigation, clean URLs, last-updated metadata, and source links back into the repo.

Current Shape

ScopeCount
Reader pages870
Core research documents866
Corpus size390k+ lines
Companies covered25
Technology domains9
Method-level SLAM library158 SLAM-method documents including overview/audit
Method-level perception files138
Safety and validation docs56
AV platform docs41
Synthesis docs12
Knowledge base docs137
Papers referenced700+
Open-source repos evaluated90+
Airport deployments documented15+

Architecture

The corpus is being organized as an end-to-end AV knowledge base: fundamentals, platform hardware, autonomy stack, runtime systems, cloud/fleet systems, safety validation, operations domains, industry intelligence, and synthesis.

Airside is used as a detailed reference ODD where the corpus has the deepest deployment evidence. Generic autonomy-stack methods, ratings, and synthesis pages should still state how ideas transfer across road AVs, warehouses, yards, ports, mines, construction sites, farms, delivery robots, and campus systems.

Start Here

NeedOpen
Navigate the whole corpusResearch Index
Get the executive viewMaster Synthesis
Start building from the researchGetting Started
Pick concrete POCsPOC Proposals
Understand readiness and riskTechnology Readiness
Prioritize gap-filling researchKnowledge Gap Backlog
Continue the research loopContinuous Research Loop
Monitor active research sourcesActive Frontier Source Registry
Compare the marketCompetitive Landscape
Read the core system architectureDesign Spec
Go deep on perception methodsMethod-Level Perception Library
Go deep on SLAM methodsMethod-Level SLAM Library
Check terms and abbreviationsGlossary
Understand how the corpus was madeMethodology

High-Leverage Reading Paths

PathBest Entry PointWhy
World models for autonomous drivingWorld Models OverviewFrames diffusion, occupancy, self-supervised occupancy flow, UniScene-style occupancy-centric generation, tokenized, JEPA, RL, and LiDAR-native approaches.
Airport airside operationsAirside Industry OverviewConnects the AV stack to pushback, turnaround, FOD, jet blast, airport data systems, and GSE.
Cross-domain deployment signals2024-2026 Autonomy Deployment IndexCompares airside, yard, warehouse, mining, delivery, and road ADS deployment evidence without treating one ODD as the default.
Safety case and certificationCertification GuidePulls together ISO 3691-4, UL 4600, SOTIF, runtime monitoring, fail-operational design, and validation.
Production deploymentDeployment PlaybookTurns research into staged rollout, shadow mode, OTA, fleet management, and operational procedures.
MLOps by scaleMLOps Scale Research Scope, MLOps Reference Architectures by Scale, MLOps Migration Checklist, MLOps Scorecards and KPIs by Scale, Data Catalog, Lineage, and Quality Ops, Experiment Tracking and Reproducibility, Model Registry and Artifact Lifecycle, Serving and Inference Operations, MLOps Platform SRE and Reliability, Pipeline Orchestration and Release Workflows, Evaluation Platforms and Replay Gates, Dataset Split and Leakage Controls, Model Monitoring and Drift Response, Site-Sliced Release Evidence, Feature and Embedding Store Operations, Offboard Labeler Registry, GPU Queueing and FinOps, Secure Artifact Attestation, Federated and Privacy-Preserving Training Policy, LLMOps and Agent Evaluation, and Map-Derived Pseudo-Label InvalidationRoutes MLOps from notebook research through repeatable prototypes, production products, fleet-scale autonomy, regulated safety-critical release, and foundation-model/platform scale, covering data-product contracts, catalog/lineage quality gates, run authority and reproducibility levels, registry identity and alias authority, serving/inference manifests, endpoint and batch-serving architecture, platform SRE/reliability, workflow state machines, evaluation manifests, replay/runtime gates, shadow/canary evidence, split/leakage firewalls, monitoring/drift response, label operations, registries, orchestration, governance, autonomy-specific map/model/calibration evidence, centralization boundaries, concrete S0-S5 architecture patterns, migration gates, release blockers, scorecards, ODD-cell release evidence, feature/vector-store controls, offboard labeler governance, GPU queueing/FinOps, secure artifact signing/SBOM/provenance/registry verification, federated/hybrid/privacy-preserving training triggers, GenAIOps/LLMOps agent evaluation, and pseudo-label invalidation.
Fleet economicsFleet TCO Business CaseTracks vehicle CAPEX, labor savings, certification costs, operator ratios, and break-even logic.
Edge hardware choicesNVIDIA Orin TechnicalGrounds model choices in compute, power, TensorRT, DLA, and sensor constraints.
Perception stackProduction Perception SystemsCompares production AV approaches and the perception patterns that transfer across road, airside, and managed-site autonomy.
Method-level perceptionPerception Method LibrarySplits BEV, sparse-query detection, occupancy, GaussianFlowOcc/GaussTR/GS-Occ3D/VOGS-CP-style Gaussian occupancy and label curation, LiDAR-camera/radar-camera fusion, dynamic Gaussian/3DGS/4DGS, point-cloud Mamba/SSM backbones, LOSC-style LiDAR pseudo-label consolidation, LiDAR MOS, scene flow, 4D radar, FMCW LiDAR, open-world occupancy/attributes including SpaCeFormer-style open-vocabulary 3D instance segmentation, robust fusion, QuantV2X/SparseCoop-style V2X compression and sparse-query cooperation, latency, and data-engine methods into single-technique research pages.
Aggregated-map semantic segmentationAggregated-Map Semantic SegmentationRoutes the full registered-map labeling pipeline: LiDAR/image inputs and modality release contracts, ML-related SLAM substrate scope, application architecture patterns for runtime maps/training exports/monitoring/digital twins/benchmarks, first-principles map-state/evaluation grounding, georeferenced source-map conditioning via OpenLiDARMap/FlexCloud-style provenance gates, source-map acceptance packages, LAMM/Uni-Mapper-style multi-session map merging plus MapEval-style source-map geometry QA before segmentation, release-oriented dataset selection including non-road urban-district benchmark bundles, GridNet-HD managed-site LiDAR-image transfer, and Point Cloud City / City-Facade / ZAHA managed-building and facade proxies, class-taxonomy design with a semantic-vs-permanence layer split, release-state-aware training objectives and map-derived training export masks that require semantic class plus release-state labels, modality-aware backbone/training/deployment-contract tradeoffs, schema-backed semantic-map/runtime contracts with map-hygiene layer digests and metrics, compact proxy/input/training selector, point-cloud SSM efficiency frontiers, LOSC-style pseudo-label consolidation, training architectures, tiling/stitching with tile release ledgers and release-state seam checks, post-processing with semantic/confidence/hygiene layer outputs and release-state-preserving smoothing, map cleaning handoff, map-hygiene ground-truth workflow, and non-road urban-district transfer.
ML-related SLAM for semantic mapsML-Related SLAM Research ScopeConnects learned registration, learned place recognition, semantic/dynamic SLAM, neural implicit and Gaussian SLAM, point-cloud removal, static-but-transient quarantine, layered removal labels, multi-session map merging, learned map priors, and aggregated-map semantic segmentation into one research architecture with an explicit handoff contract for what may steer publication versus QA or training.
LiDAR artifact removalLiDAR Artifact Removal TechniquesConnects LIORNet, learned denoisers, classical outlier filters, weather artifacts, ghost/multipath behavior, map cleaning, datasets, and safety validation.
Dynamic and static object removalLiDAR Map Cleaning and Dynamic RemovalConnects ERASOR, Removert, MapCleaner, ERASOR++, 4dNDF, FreeDOM, BeautyMap, Raymoval, STATIC-LIO, MOVES, detector-based potentially dynamic object removal, dynamic residual cleanup, stationary-person hard exclusion, static-but-transient quarantine, reason-coded release decisions and removal sidecars, release-state benchmark labels, static-but-wrong map exclusion, Uni-Mapper dynamic-aware heterogeneous-LiDAR map merging, LAMM large-scale multi-session point-cloud map merging, MapEval post-cleaning geometry QA, RTMap/DUFOMap, LT-mapper/Khronos, lifelong map version control, MOS/scene-flow methods, moved-object datasets, and false-deletion validation.
Perception coverage gapsPerception Coverage AuditTracks missing first-class perception pages across BEV, occupancy, Gaussian/3DGS, LiDAR/radar/thermal, open-world/OOD, V2X, robustness, and benchmarks.
Localization and mappingMapping and LocalizationCovers HD maps, LiDAR SLAM, map-prior georeferenced construction, multi-session point-cloud map merging, point-cloud map-quality evaluation, map-free driving, map maintenance, infrastructure-aided localization, and occupancy grids.
Photoreal city-scale 4D reconstructionPhotoreal city-scale 4D reconstructionLinks Gaussian SLAM, VGGT/feed-forward reconstruction, dynamic 4D Gaussian/NeRF methods, and digital-twin simulation coverage.
Method-level 3D SLAMSLAM Library OverviewBreaks classical, LiDAR including RKO-LIO sensor-agnostic LIO, LIVO, visual, dense, neural, Gaussian, radar, learned 4D radar odometry, radar place-recognition descriptors, radar RIO correspondence/uncertainty, Doppler radar-LiDAR bridge SLAM, raw GNSS factor fusion, wheel/vehicle-motion factors, radar-GNSS/visual mapping, multi-sensor, SLAM Toolbox, NDT variants, static-map lifecycle/removal, heterogeneous map-merging, LAMM multi-session point-cloud map merging, MapEval map-quality evaluation, and current SLAM benchmark coverage into focused method files.
GLIM and GTSAM pipelineGLIM and GTSAM Pipeline HubMaps GLIM pipeline stages to GTSAM objects, Bayes-tree updates, Hessians, sparse elimination, marginalization, and the supporting KB math pages.
SLAM coverage gapsSLAM Coverage AuditTracks promoted and remaining first-class SLAM pages, including May 2026 sweeps across LIO, LIVO, raw GNSS factor fusion, wheel/vehicle-motion factors, 4D radar, Gaussian/foundation SLAM, backends, collaborative SLAM, alternative sensors, and benchmarks.
First-principles estimator mathNonlinear Solver Diagnostics CrosswalkRoutes estimator failures across residuals, Jacobians, scaling, damping, rank, covariance, constrained KKT/QP/SQP mechanics, robust-loss covariance consistency, fiducial/corner localization, two-view epipolar/homography verification, optical/scene-flow motion fields, and sparse backend choices, with links back into probability, optimization, numerical linear algebra, geometry, and state estimation foundations.
Machine learning foundationsMachine Learning FoundationsStarts from linear models and gradients through CNN/RNN/transformer/SSM foundations, self-supervision, world models, AV data-evaluation contracts, calibration, evaluation, and deployment review.
Control and decision foundationsControl FoundationsStarts the foundations path for closed-loop tracking, vehicle dynamics, MPC/iLQR, constraints, MDP/POMDP decision models, safety filters, and planner-controller review.
Sensor and estimation fundamentalsSensor FoundationsStarts the sensor-model foundation path, with supporting links into geometry, thermal IR radiometry, ultrasonic proximity sensing, state estimation, signal processing, timing, calibration, and wheel odometry.
Sensor readiness before algorithmsSensor-to-Algorithm Readiness ContractConsolidates calibration, synchronization, preprocessing, health, provenance, and fail-closed gates before perception, fusion, SLAM, tracking, occupancy, mapping, or planning consumes sensor-derived inputs.
Perception validation datasetsFOD and Airport Apron Detection DatasetsConnects MUSES, DSERT-RoLL, CMHT, EmbodiedScan/MMScan, GridNet-HD, non-road district benchmark bundles, STU 3D anomaly segmentation, RCP-Bench, TruckV2X, V2X datasets, sensor-corruption benchmarks, open-world/OOD anomaly segmentation, FOD datasets, Airport-FOD3S synthetic data-engine routing, synthetic FOD validation, FOD validation, public-proxy boundary caveats, and knowledge-base evaluation protocols.
End-to-end architecture gapsKnowledge Gap BacklogTracks P0/P1/P2 missing research files across fundamentals, platform, autonomy, runtime/cloud, safety, operations, and industry intelligence.

Corpus Map

SectionDocsStart AtWhat It Holds
00-start-here/4Reading GuideReader entry points and orientation material.
10-knowledge-base/137Probability and Statistics FoundationsFirst-principles technical notes: probability/statistics, optimization, constrained KKT/QP/SQP solver mechanics, numerical linear algebra, geometry, mapping, state estimation, sensor likelihoods, thermal IR radiometry, ultrasonic proximity sensing, signal processing, controls, robotics, ML including AV data-evaluation contracts, calibration, timing, continuous-time trajectories, robust covariance consistency, fiducial/corner localization, two-view epipolar/homography verification, optical/scene-flow motion fields, and detection/tracking evidence.
20-av-platform/41NVIDIA Orin TechnicalCompute, sensors, sensor-to-algorithm readiness, connectivity, drive-by-wire, power, diagnostics, ruggedization, and edge-cloud architecture.
30-autonomy-stack/459World Models OverviewWorld models, perception, method-level perception, planning, localization, infrastructure-aided localization, SLAM, simulation, VLA/VLM, E2E driving, and multi-agent systems.
40-runtime-systems/23Production ML DeploymentML deployment, ROS/Autoware, observability, teleoperation, software operations, and vehicle-side data logging.
50-cloud-fleet/52Cloud Backend InfrastructureData engines, fleet data loops, MLOps, OTA/SUMS, observability, map operations, data governance, perception/SLAM reliability telemetry, and fleet management.
60-safety-validation/56Certification GuideSafety case, standards, runtime assurance, verification, validation, robustness, cybersecurity, incident reporting, reliability evidence, and evidence traceability.
70-operations-domains/27Airside Industry OverviewAirside, warehouse, yard, port, mining, agriculture, construction, road AV, delivery robot, deployment, business-case, and safety operations.
80-industry-intel/62Company IndexAV, airside, simulation, teleoperation, autonomy company profiles, market intelligence, and regulations.
90-synthesis/12Master SynthesisExecutive synthesis, POCs, readiness, risk, decision framework, architecture, gap backlog, continuous research loop, and active frontier source registry.

Domain Snapshot

TechnologyDocs
World models18
Perception212
Method-level perception library138
Planning16
Localization and mapping185
Method-level SLAM library158 SLAM-method documents including overview/audit
Simulation8
VLA / VLM7
Multi-agent and V2X7
Robustness validation files9
E2E driving6
OperationsDocs
Safety and validation56
Deployment13
Airside operations11
Cross-domain operations9
Teleoperation1
AV PlatformDocs
Compute9
Sensors19
Networking/connectivity6
Drive-by-wire2
Power/electrical1
Diagnostics2
Ruggedization1
Thermal1

Reader Notes

  • The static reader is generated from this repository with VitePress and deployed through GitHub Pages.
  • README.md becomes the site home page.
  • INDEX.md is served as /INDEX/ in the reader to avoid a Windows case-insensitive output collision with the homepage.
  • Research content is source-of-truth Markdown; the generated site is just a browser-friendly layer over the same files.
  • Internal implementation notes under docs/superpowers/, .claude/, and .superpowers/ are excluded from the static reader.

Public research notes collected from public sources.