Research Index
Quick Navigation by Topic
If you need to know about...
A specific company
| Company | Primary | Also mentioned in |
|---|---|---|
| Waymo | 80-industry-intel/companies/waymo/ (5 docs) | 30-autonomy-stack/end-to-end-driving/company-approaches.md, 60-safety-validation/safety-case/safety-incidents-lessons.md, 50-cloud-fleet/ota/ota-fleet-management.md, 50-cloud-fleet/fleet-management/fleet-management-dispatch.md, 30-autonomy-stack/perception/overview/production-perception-systems.md |
| Tesla | 80-industry-intel/companies/tesla/ (4 docs) | 30-autonomy-stack/end-to-end-driving/company-approaches.md, 60-safety-validation/safety-case/safety-incidents-lessons.md, 50-cloud-fleet/ota/ota-fleet-management.md, 30-autonomy-stack/perception/overview/production-perception-systems.md |
| comma.ai | 80-industry-intel/companies/comma-ai/ (2 docs) | 30-autonomy-stack/world-models/opensource-implementations.md, 60-safety-validation/verification-validation/shadow-mode.md, 40-runtime-systems/ml-deployment/opensource-ecosystem.md |
| UISEE | 80-industry-intel/companies/uisee/tech-stack.md | 80-industry-intel/companies/changi-programme/, 70-operations-domains/airside/operations/industry-overview.md |
| TractEasy/EasyMile | 80-industry-intel/companies/tracteasy/ (2 docs) | 60-safety-validation/standards-certification/iso-3691-4-deep-dive.md, 70-operations-domains/airside/operations/industry-overview.md |
| Wayve | 80-industry-intel/companies/wayve/ (4 docs) | 30-autonomy-stack/end-to-end-driving/company-approaches.md, 30-autonomy-stack/world-models/overview.md |
| AeroVect | 80-industry-intel/companies/aerovect/tech-stack.md | 70-operations-domains/airside/operations/industry-overview.md |
| Assaia | 80-industry-intel/companies/assaia/tech-stack.md | 80-industry-intel/companies/moonware/halo-operations.md |
| Fernride | 80-industry-intel/companies/fernride/tech-stack.md | 40-runtime-systems/monitoring-observability/teleoperation-systems.md |
| Applied Intuition | 80-industry-intel/companies/applied-intuition/tech-stack.md | 30-autonomy-stack/simulation/airport-digital-twins.md |
World models
| Topic | Primary | Supporting |
|---|---|---|
| What are world models | 30-autonomy-stack/world-models/overview.md | 90-synthesis/master/master-synthesis.md |
| World-model first principles | 10-knowledge-base/machine-learning/world-models-first-principles.md | Latent state, transition models, observation/reward heads, Dreamer/PlaNet, tokenized, diffusion, and JEPA branches |
| Diffusion-based | 30-autonomy-stack/world-models/diffusion-world-models.md | 10-knowledge-base/machine-learning/diffusion-models.md |
| Occupancy-based | 30-autonomy-stack/world-models/occupancy-world-models.md | 30-autonomy-stack/world-models/occupancy-networks-comparison.md (20 methods) |
| Tokenized / JEPA | 30-autonomy-stack/world-models/tokenized-and-jepa.md | 10-knowledge-base/machine-learning/vqvae-tokenization.md, 10-knowledge-base/machine-learning/jepa-latent-predictive-learning.md |
| RL with world models | 30-autonomy-stack/world-models/rl-with-world-models.md | 30-autonomy-stack/world-models/dreamer-world-model-rl.md |
| OccWorld setup | 30-autonomy-stack/world-models/occworld-implementation.md | 30-autonomy-stack/world-models/occupancy-networks-comparison.md |
| Open-source repos | 30-autonomy-stack/world-models/opensource-implementations.md | 21 repos rated |
| Cutting edge 2026 | 30-autonomy-stack/world-models/cutting-edge-2026.md | Latest papers and SOTA |
| Occupancy on Orin | 30-autonomy-stack/world-models/occupancy-deployment-orin.md | FlashOcc TensorRT, nvblox, LiDAR voxelization, multi-resolution grids |
| LiDAR-native world models | 30-autonomy-stack/world-models/lidar-native-world-models.md | Copilot4D, UnO, LidarDM, LiDARCrafter, 4D occupancy forecasting, point cloud prediction, AD-L-JEPA, self-supervised training, Orin deployment caveats |
| Occupancy flow & 4D scenes | 30-autonomy-stack/world-models/occupancy-flow-4d-scenes.md | Scene flow (ZeroFlow 0.028m EPE, DeFlow SOTA), 4D occupancy forecasting (UnO, OccSora, Cam4DOcc), dynamic 3D Gaussians, K-Planes 10900x compression, flow-guided Frenet planning, Mamba temporal, Orin 26-40ms FP16, $6-11K training |
| Self-supervised occupancy flow | 30-autonomy-stack/world-models/self-supervised-occupancy-flow.md | Let Occ Flow, SelfOccFlow, static/dynamic field decomposition, self-supervised 3D occupancy-flow training, and label-cost reduction for dynamic scenes |
| Occupancy-centric scene generation | 30-autonomy-stack/world-models/uniscene-occupancy-centric-generation.md | UniScene-style semantic occupancy as the shared representation for generated video, LiDAR, and inspectable synthetic-data supervision |
| Scene flow for removal | 30-autonomy-stack/world-models/scene-flow-for-dynamic-object-removal.md | Connects LiDAR scene flow, MOS, occupancy flow, static-map cleaning, flow-to-map hygiene decisions, and planner-facing dynamic-object evidence |
| Scene-flow benchmarks | 30-autonomy-stack/world-models/scene-flow-datasets-benchmarks.md | FlyingThings3D, KITTI Scene Flow, Argoverse 2 flow, Waymo flow, ZeroFlow/DeFlow evaluation, and removal-oriented metrics |
Machine learning foundations
| Topic | Primary | Supporting |
|---|---|---|
| ML foundation ladder | 10-knowledge-base/machine-learning/overview.md | Reading path from perceptron and logits through backprop, optimization, CNNs, RNNs, transformers, Mamba, JEPA, and world models |
| Linear and probabilistic classifiers | 10-knowledge-base/machine-learning/perceptron-linear-classifiers.md | 10-knowledge-base/machine-learning/logistic-softmax-cross-entropy.md |
| Training mechanics | 10-knowledge-base/machine-learning/backprop-computational-graphs-autodiff.md | 10-knowledge-base/machine-learning/optimization-training-dynamics.md, 10-knowledge-base/machine-learning/initialization-normalization-regularization.md |
| Spatial and temporal neural networks | 10-knowledge-base/machine-learning/convolutional-neural-networks.md | 10-knowledge-base/machine-learning/recurrent-neural-networks-lstm-gru.md, 10-knowledge-base/machine-learning/sequence-models-rnn-ssm-attention-first-principles.md |
| Transformer and foundation models | 10-knowledge-base/machine-learning/attention-transformers-first-principles.md | 10-knowledge-base/machine-learning/vision-transformers-first-principles.md, 10-knowledge-base/machine-learning/foundation-model-training-first-principles.md |
| Self-supervised and predictive learning | 10-knowledge-base/machine-learning/self-supervised-learning-first-principles.md | 10-knowledge-base/machine-learning/jepa-latent-predictive-learning.md, 10-knowledge-base/machine-learning/world-models-first-principles.md |
| Representation objectives | 10-knowledge-base/machine-learning/contrastive-learning-infonsce-first-principles.md | 10-knowledge-base/machine-learning/masked-modeling-first-principles.md, 10-knowledge-base/machine-learning/energy-based-models-first-principles.md, 10-knowledge-base/machine-learning/autoencoders-vae-and-latent-variable-models-first-principles.md |
| Sequence, tokens, and generators | 10-knowledge-base/machine-learning/state-space-models-s4-mamba-first-principles.md | 10-knowledge-base/machine-learning/tokenization-and-discretization-first-principles.md, 10-knowledge-base/machine-learning/positional-encodings-and-coordinate-tokenization-first-principles.md, 10-knowledge-base/machine-learning/diffusion-score-flow-samplers-first-principles.md |
| Evaluation and objective design | 10-knowledge-base/machine-learning/av-data-evaluation-fundamentals.md, 10-knowledge-base/machine-learning/evaluation-calibration-and-data-leakage-first-principles.md | 10-knowledge-base/machine-learning/multi-task-losses-and-objectives-first-principles.md, 10-knowledge-base/machine-learning/world-model-evaluation-and-planning-objectives-first-principles.md |
Perception
| Topic | Primary | Supporting |
|---|---|---|
| BEV encoding | 30-autonomy-stack/perception/overview/bev-encoding.md | 10-knowledge-base/geometry-3d/pointpillars.md |
| Open-vocab detection | 30-autonomy-stack/perception/overview/open-vocab-detection.md | YOLO-World, Grounding DINO |
| DINOv2 for driving | 30-autonomy-stack/perception/overview/dinov2-foundation-models-driving.md | LoRA, adapter integration |
| CenterPoint/OpenPCDet | 30-autonomy-stack/perception/overview/openpcdet-centerpoint.md | 20-av-platform/compute/tensorrt-deployment-guide.md |
| Production systems | 30-autonomy-stack/perception/overview/production-perception-systems.md | Waymo/Tesla/comma sensor suites |
| Perception method library | 30-autonomy-stack/perception/methods/overview.md | 138 method-library files, including 137 atomic method pages plus the overview, across camera BEV, sparse-query detection, end-to-end driving, occupancy/free-space, Gaussian occupancy and label curation including VOGS-CP collaborative Gaussian occupancy, LiDAR-camera/radar-camera fusion, dynamic Gaussian/3DGS/4DGS, 3D segmentation backbones including point-cloud Mamba/SSM backbones, LOSC open-vocabulary LiDAR label consolidation, LiDAR MOS, scene flow, LiDAR denoising/removal, radar/4D radar, event/FMCW, open-world/OOD, open-vocabulary attributes including SpaCeFormer-style open-vocabulary 3D instance segmentation, robust fusion, V2X compression and sparse-query cooperation including QuantV2X and SparseCoop, latency, and data-engine evaluation |
| LiDAR artifact removal | 30-autonomy-stack/perception/overview/lidar-artifact-removal-techniques.md | LIORNet, LiSnowNet, SLiDE, TripleMixer, classical filters, weather artifacts, ghost/multipath behavior, dynamic-map cleaning, and validation |
| Weather robustness datasets | 30-autonomy-stack/perception/datasets-benchmarks/weather-robustness-datasets.md | WADS, CADC/CADC+, SemanticSTF, REHEARSE-3D, RainSense, SemanticSpray, RADIATE, DSERT-RoLL, CMHT, and Seeing Through Fog/DENSE |
| Moving/static separation datasets | 30-autonomy-stack/perception/datasets-benchmarks/moving-static-separation-mos-datasets.md | SemanticKITTI-MOS, HeLiMOS, 4DMOS-style labels, moving/static taxonomy, and map-cleaning evaluation fit |
| Occupancy-flow benchmarks | 30-autonomy-stack/perception/datasets-benchmarks/occupancy-flow-and-4d-occupancy-benchmarks.md | Cam4DOcc, OpenOccupancy, Occ3D/OpenScene, UniOcc, nuCraft, and 4D occupancy metrics for flow/removal systems |
| Large-scale 3D segmentation benchmarks | 30-autonomy-stack/perception/datasets-benchmarks/large-scale-3d-segmentation-benchmarks.md | SemanticKITTI single/multi-scan, SemanticTHAB, Semantic3D, Paris-Lille-3D/NPM3D, Toronto-3D, KITTI-360, DALES, GridNet-HD, ECLAIR, YUTO Semantic, S.MID, OpenTrench3D, MLDAS, USCILab3D, Industrial3D, Point Cloud City / Open3D-ML PCC, City-Facade, ZAHA, SensatUrban, WHU-Urban3D, WHU-Railway3D, CUS3D, SUM Parts, GOOSE-Ex, STPLS3D — splits, metrics, test servers, label formats, licensing, release-oriented dataset-selection protocol, non-road district benchmark bundles, and urban/non-road/utility-infrastructure/managed-site/facade proxy fit for point-cloud, LiDAR-image, and mesh semantic segmentation |
| GridNet-HD utility segmentation | 30-autonomy-stack/perception/datasets-benchmarks/gridnet-hd-power-line-lidar-image-segmentation.md | 2026 LiDAR-image benchmark for overhead electrical infrastructure: 36 zones, 7,694 images, 2.45B LiDAR points, 11 evaluated semantic groups, hidden-label leaderboard, SPT/ImageVote/late-fusion baselines, managed-site transfer protocol, and thin-class stress testing for aggregated-map segmentation |
| Adverse/OOD/FOD/V2X benchmarks | 30-autonomy-stack/perception/datasets-benchmarks/muses-multisensor-adverse-semantic-perception.md, 30-autonomy-stack/perception/datasets-benchmarks/dsert-roll.md, 30-autonomy-stack/perception/datasets-benchmarks/cmht-autonomous-dataset.md, 30-autonomy-stack/perception/datasets-benchmarks/sensor-corruption-robustness-benchmarks.md, 30-autonomy-stack/perception/datasets-benchmarks/open-world-ood-anomaly-segmentation-benchmarks.md, 30-autonomy-stack/perception/datasets-benchmarks/stu-3d-lidar-anomaly-segmentation.md, 30-autonomy-stack/perception/datasets-benchmarks/fod-and-airport-apron-detection-datasets.md, 30-autonomy-stack/perception/datasets-benchmarks/airside-fod-synthetic-multimodal-benchmarks.md, 50-cloud-fleet/data-platform/airport-fod3s-synthetic-data.md, 30-autonomy-stack/perception/datasets-benchmarks/rcp-bench-cooperative-corruption-robustness.md, 30-autonomy-stack/perception/datasets-benchmarks/v2x-large-range-sequential-datasets.md, 30-autonomy-stack/perception/datasets-benchmarks/truckv2x-truck-centered-cooperative-perception.md | MUSES, DSERT-RoLL, CMHT, Robo3D/MultiCorrupt-style corruption tests, STU 3D anomaly segmentation, SegmentMeIfYouCan/OpenAD-style anomaly segmentation, airport FOD and synthetic multimodal FOD benchmark framing with public-proxy caveats, Airport-FOD3S data-engine workflow, cooperative corruption robustness, large-range V2X datasets, and truck-centered cooperative perception |
| Embodied 3D perception benchmarks | 30-autonomy-stack/perception/datasets-benchmarks/embodiedscan-mmscan-embodied-3d-benchmarks.md | EmbodiedScan and MMScan for egocentric RGB-D 3D perception, semantic occupancy, visual grounding, 3D QA, and VLM/VLA spatial-grounding evaluation |
| Perception coverage audit | 30-autonomy-stack/perception/overview/coverage-audit-2026.md | May 2026 multi-agent sweeps across camera BEV/occupancy, LiDAR MOS, 4D radar, open-world/OOD, V2X, robust fusion, deployment validation, and benchmarks |
| Sensor fusion | 30-autonomy-stack/perception/overview/sensor-fusion-architectures.md | BEVFusion, masked modality training |
| Camera-LiDAR fusion interfaces | 30-autonomy-stack/perception/overview/camera-lidar-fusion-interfaces.md | Projection, BEV/query/voxel/late fusion, offline map colorization, LiDAR-only vs colorized vs distillation vs image-dependent release contracts, and projection QA evidence |
| Infrastructure cooperative perception | 30-autonomy-stack/perception/overview/infrastructure-cooperative-perception.md | V2I fusion, fixed sensors, DAIR-V2X, QuantV2X/SparseCoop/VOGS-CP communication primitives, TruckV2X, airport existing systems |
| LiDAR foundation models | 30-autonomy-stack/perception/overview/lidar-foundation-models.md | PTv3, Sonata, ScaLR, PointLoRA, 50-80% data savings |
| LiDAR semantic segmentation | 30-autonomy-stack/perception/overview/lidar-semantic-segmentation.md | Cylinder3D, FlatFormer, PTv3, ALPINE panoptic, airside 18-class taxonomy |
| Aggregated-map semantic segmentation | 30-autonomy-stack/perception/overview/aggregated-map-semantic-segmentation.md | End-to-end pipeline for segmenting registered multi-scan LiDAR maps: tiling/stitching with tile release ledgers, urban/non-road/utility-infrastructure/managed-site/facade proxy datasets including GridNet-HD, Point Cloud City / Open3D-ML PCC, City-Facade, and ZAHA, ML-related SLAM substrate scope, source-map acceptance packages, georeferenced source-map conditioning via OpenLiDARMap/FlexCloud-style provenance gates, LAMM/Uni-Mapper-style multi-session map merging plus MapEval source-map geometry QA before segmentation, permanence decision layer separating motion, semantic mobility, persistence, operations, and map eligibility, modality-aware sparse-conv/KPConv/RandLA/SPT/PTv3/Sonata/projection/SSM architecture and training-route comparison, compact proxy/input/training selector, LiDAR±image distillation and modality release-contract lanes, schema-backed semantic-map/runtime contracts, LOSC pseudo-label consolidation, post-processing semantic/confidence/hygiene layer outputs, map-hygiene ground-truth gates, mesh/digital-twin transfer, auto-label flywheel |
| ML-related SLAM for semantic maps | 30-autonomy-stack/localization-mapping/overview/ml-related-slam-research-scope.md | Research scope linking learned registration, learned place recognition, semantic/dynamic SLAM, neural implicit and Gaussian SLAM, point-cloud removal, layered removal labels, motion/permanence/map-eligibility separation, static-but-transient quarantine, map priors, multi-session map merging, downstream aggregated-map segmentation, and a handoff contract for what learned evidence may affect in release |
| Aggregated-map segmentation companions | 30-autonomy-stack/perception/overview/3d-segmentation-class-taxonomy-design.md, 30-autonomy-stack/perception/overview/3d-segmentation-training-paradigms.md, 30-autonomy-stack/perception/overview/large-scale-3d-segmentation-tiling-and-throughput.md, 30-autonomy-stack/perception/overview/segmentation-post-processing-label-refinement.md, 30-autonomy-stack/perception/overview/static-but-transient-point-removal.md, 30-autonomy-stack/perception/datasets-benchmarks/moving-static-separation-mos-datasets.md, 30-autonomy-stack/perception/methods/point-cloud-mamba-ssm-backbones.md, 30-autonomy-stack/perception/methods/losc.md, 30-autonomy-stack/localization-mapping/maps/airside-map-hygiene-ground-truth-protocol.md, 60-safety-validation/verification-validation/airside-map-hygiene-ground-truth-protocol.md, 30-autonomy-stack/localization-mapping/slam-methods/potentially-dynamic-object-removal-ground-projection.md, 30-autonomy-stack/localization-mapping/slam-methods/uni-mapper-dynamic-aware-lidar-map-merging.md, 30-autonomy-stack/localization-mapping/slam-methods/lamm-multi-session-point-cloud-map-merging.md, 30-autonomy-stack/localization-mapping/slam-methods/mapeval-point-cloud-map-quality-evaluation.md | Class-taxonomy design, training architecture comparison with map-derived label eligibility masks, map-scale tiling/throughput with tile release ledgers and release-state seam confusion, label-refinement/post-processing with semantic/confidence/hygiene output layers and release-state-preserving smoothing guardrails, stationary-transient object removal, release-state benchmark labels, point-cloud SSM/Mamba backbone candidates, LOSC-style pseudo-label consolidation, canonical map-hygiene ground-truth and V&V workflow, detector-based movable-object quarantine, dynamic-aware heterogeneous-LiDAR map merging, large-scale multi-session map merging, and source-map geometry QA for registered LiDAR/RGB map products |
| Semantic-map release contracts | schemas/semantic-map-manifest.schema.json, schemas/runtime-map-contract.schema.json, examples/map-contracts/, tools/map-contracts/validate.mjs | JSON Schema gates and examples for semantic-map manifests, runtime map contracts, prior-input provenance, map-hygiene layer digests, map-hygiene metric vectors, training-export release-state eligibility, artifact-set compatibility, and CI validation before a labeled aggregated map is published or consumed |
| Model compression & edge | 30-autonomy-stack/perception/overview/model-compression-edge-deployment.md | PTQ/QAT, distillation, pruning, TensorRT, ModelOpt, Orin recipes |
| Multi-object tracking | 30-autonomy-stack/perception/overview/multi-object-tracking.md | CenterPoint tracker, SimpleTrack, MCTrack, HOTA, airside Re-ID |
| Camera fallback perception | 30-autonomy-stack/perception/overview/camera-fallback-perception.md | Degraded mode when LiDAR fails: DepthAnything v2, stereo depth, BEVFormer-Tiny, confidence calibration, speed reduction |
| Collaborative fleet perception | 30-autonomy-stack/perception/overview/collaborative-fleet-perception.md | V2V cooperative sensing, Where2comm bandwidth selection, CoBEVT/CoBEVFlow temporal fusion, HEAL heterogeneous agents, QuantV2X compression, SparseCoop sparse queries, VOGS-CP collaborative Gaussian occupancy, TruckV2X heavy-vehicle proxy, fleet occupancy map, collective FOD detection, 5G deployment |
| V2X protocols & airside messages | 30-autonomy-stack/multi-agent-v2x/v2x-protocols-airside.md | C-V2X vs DSRC (5G NR V2X preferred), ETSI ITS (CAM/DENM/CPM/MCM), 8 airside-specific messages (APA, SOS, GTA, DZN, EVP, RIP, FDA, JBW), protobuf specs, A-CDM/A-SMGCS/ADS-B bridge, PKI security, bandwidth planning (123 Mbps/50 vehicles), default-deny runway clearance, $270-450K full capability |
| Fleet task allocation & scheduling | 30-autonomy-stack/multi-agent-v2x/fleet-task-allocation-scheduling.md | MRTA MT-SR-TA formulation, MILP/CP-SAT (OR-Tools optimal in 10-60s for 200 vehicles), Hungarian O(n³) single-assignment, CBBA decentralized auction (95% optimal, <100ms), SSI real-time auction, A-CDM predictive scheduling (ELDT→pre-positioning, 60-75% delay reduction), online reactive scheduling (event-driven rescheduling, 85% stability), RL dispatch policy (<1ms inference), charging-aware scheduling, multi-objective (tardiness+energy+safety), priority-based task shedding, $42-67K/15-17 weeks |
| Ramp traffic conflict & deadlock prevention | 30-autonomy-stack/multi-agent-v2x/ramp-traffic-conflict-deadlock-prevention.md | Zone-capacity graph from Lanelet2, reservation-based traffic management, wait-die deadlock prevention (guarantees no circular wait), 9-level priority conflict resolution, stand turnaround sequencing, V2X decentralized fallback, token mutex for single-lane zones, MAPF (CBS/ECBS for offline, PIBT for real-time), livelock detection/resolution, capacity-constrained routing, dispatch-traffic integration, $50-75K/17 weeks |
| Self-supervised pre-training | 30-autonomy-stack/perception/overview/self-supervised-pretraining-driving.md | Contrastive (SLidR, ScaLR), MAE (Voxel-MAE, GD-MAE, BEV-MAE), JEPA (AD-L-JEPA, V-JEPA 2), DINOv2, multi-modal pre-training, LoRA fine-tuning, 50-80% label reduction, airside curriculum strategy |
| 3DGS for perception & mapping | 30-autonomy-stack/perception/overview/gaussian-splatting-driving.md | GaussianFormer/GaussianOcc, SplatAD, streaming Gaussian occupancy, SplaTAM/MonoGS/Splat-SLAM/S3PO-GS, LiDAR-Gaussian fusion, dynamic object tracking, semantic Gaussians, FOD detection, aircraft proximity, Orin deployment notes |
| Dynamic Gaussian/neural-field perception | 30-autonomy-stack/perception/methods/drivinggaussian.md, 30-autonomy-stack/perception/methods/hugs-urban-gaussians.md, 30-autonomy-stack/perception/methods/splatflow.md, 30-autonomy-stack/perception/methods/distillnerf.md | Dynamic 3DGS/4DGS, holistic urban Gaussians, self-supervised Gaussian motion flow, and NeRF-to-occupancy distillation for perception and simulation reuse |
| Photoreal city-scale 4D reconstruction | 30-autonomy-stack/localization-mapping/overview/photoreal-city-scale-4d-reconstruction.md, 10-knowledge-base/geometry-3d/feed-forward-3d-reconstruction-and-splatting.md, 10-knowledge-base/mapping/dynamic-4d-neural-gaussian-reconstruction.md | Cross-section hub and first-principles pages for Gaussian-LIC/LIC2, RMGS-SLAM, VGGT, AnySplat, pixelSplat, Street Gaussians, OmniRe, S3Gaussian, EmerNeRF, OG-Gaussian, PVG, and DrivingGaussian |
| 4D radar-camera, radar-LiDAR, and FMCW perception | 30-autonomy-stack/perception/methods/cvfusion.md, 30-autonomy-stack/perception/methods/4d-radar-camera-occupancy.md, 30-autonomy-stack/perception/methods/adverse-weather-radar-lidar-3d-detection.md, 30-autonomy-stack/perception/methods/robucdet.md, 30-autonomy-stack/perception/methods/samfusion.md, 30-autonomy-stack/perception/methods/pod-fmcw-lidar-predictive-detection.md | Cross-view radar-camera detection, radar-camera semantic occupancy, radar-LiDAR adverse-weather detection, robust radar-camera BEV, sensor-adaptive multimodal fusion, and FMCW LiDAR velocity-aware predictive detection |
| Occupancy fusion and open-world occupancy | 30-autonomy-stack/perception/methods/lidar-camera-occupancy-fusion.md, 30-autonomy-stack/perception/methods/dynamic-occupancy-freespace.md, 30-autonomy-stack/perception/methods/spatiotemporal-memory-occupancy-flow.md, 30-autonomy-stack/perception/methods/open-vocabulary-panoptic-occupancy.md, 30-autonomy-stack/perception/methods/ovad-ovoda-open-vocab-3d-attributes.md, 30-autonomy-stack/perception/methods/spaceformer.md, 30-autonomy-stack/perception/methods/losc.md | LiDAR-camera semantic occupancy fusion, dynamic/free-space occupancy, temporal occupancy memory, language/panoptic occupancy, open-vocabulary 3D attributes, proposal-free open-vocabulary 3D instance segmentation, and open-vocabulary LiDAR pseudo-label consolidation for state-rich object semantics |
| 3D point cloud segmentation backbones | 30-autonomy-stack/perception/methods/minkowskinet.md, 30-autonomy-stack/perception/methods/kpconv.md, 30-autonomy-stack/perception/methods/randla-net.md, 30-autonomy-stack/perception/methods/cylinder3d.md, 30-autonomy-stack/perception/methods/waffleiron.md, 30-autonomy-stack/perception/methods/point-transformer-v3.md, 30-autonomy-stack/perception/methods/octformer.md, 30-autonomy-stack/perception/methods/superpoint-transformer.md, 30-autonomy-stack/perception/methods/point-cloud-mamba-ssm-backbones.md | Deep-dive architecture pages for the model families behind LiDAR semantic segmentation and aggregated-map labeling: sparse-voxel conv, point conv, efficient large-scale point networks, cylindrical voxels, projection-based dense 2D conv, serialized/octree transformers, superpoint-graph segmentation, and the state-space/SSM efficiency frontier |
| Sparse-query and end-to-end driving | 30-autonomy-stack/perception/methods/sparsebev.md, 30-autonomy-stack/perception/methods/sparse4d.md, 30-autonomy-stack/perception/methods/detr4d.md, 30-autonomy-stack/perception/methods/foresight.md, 30-autonomy-stack/perception/methods/sparsedrive.md, 30-autonomy-stack/perception/methods/diffusiondrive.md, 30-autonomy-stack/perception/methods/sam4d.md, 30-autonomy-stack/perception/methods/open3dtrack-open-vocab-3d-tracking.md | Sparse object queries, temporal camera 3D detection, sparse end-to-end perception-planning stacks, diffusion planning policies, open-vocabulary 4D segmentation, and open-vocabulary 3D tracking |
| Uncertainty quantification | 30-autonomy-stack/perception/overview/uncertainty-quantification-calibration.md | Epistemic/aleatoric decomposition, MC-Dropout (T=3, 21.5ms), deep ensembles (M=5, 0.93 AUROC), evidential deep learning (single pass, 7.5ms), conformal prediction (99% coverage guarantee), temperature scaling (ECE 0.03), LiDAR range-dependent uncertainty, multi-LiDAR fusion (65% reduction), teleop trigger criteria |
| Multi-task unified perception | 30-autonomy-stack/perception/overview/multi-task-unified-perception.md | UniAD (CVPR 2023 Best Paper), SparseDrive (3x faster), VAD-Tiny (80ms Orin), StreamPETR, shared-backbone multi-head (14.8ms on Orin, 56% savings), task interference/PCGrad, uncertainty-weighted loss, incremental deployment, 14-class airside segmentation |
| Night operations & thermal fusion | 30-autonomy-stack/perception/overview/night-operations-thermal-fusion.md | LiDAR-primary + thermal-augmented architecture, YOLO-Thermal INT8 (6-8ms Orin), asymmetric late fusion (+8-10ms), hi-vis paradox solved (84-88% camera AEB failure → 85-92% thermal AP), heated-target calibration (<0.5deg), jet blast/fuel spill thermal detection, night ODD (subset of daytime), DINOv2 LoRA thermal adapter, 22.8-25.8ms total pipeline (38-44 Hz), $6,700-22,600/vehicle |
| Streaming temporal perception | 30-autonomy-stack/perception/overview/streaming-temporal-perception.md | StreamPETR (+6-8% NDS, <3ms overhead, implicit tracking), Sparse4D v3 (71.9% NDS SOTA), multi-sweep LiDAR accumulation (3-sweep: +2.5% mAP, +1.4ms), latency compensation (ASAP/LASP), temporal filtering eliminates transient noise (de-icing spray, jet blast shimmer), extended airside track persistence (10-30s for GSE occlusion), video backbones vs query propagation, turnaround phase detection, $38K/13 weeks |
| Active perception & sensor scheduling | 30-autonomy-stack/perception/overview/active-perception-sensor-scheduling.md | Context-aware model switching (35-45% compute reduction), information-theoretic sensor selection (entropy-based attention), foveated LiDAR (89% voxel reduction), multi-LiDAR scheduling (3-4 of 8 LiDARs full at any time), early exit networks (48% average compute), risk-aware allocation (safety-critical always first), planner-guided attention, 30-36% power savings for electric GSE, $25-40K/10 weeks |
Method-level SLAM
| Topic | Primary | Supporting |
|---|---|---|
| SLAM method library | 30-autonomy-stack/localization-mapping/slam-methods/overview.md | 158 SLAM-method documents including overview/audit pages covering classical, LiDAR including RKO-LIO sensor-agnostic LIO, LIVO, visual, RGB-D, neural, Gaussian, radar, learned 4D radar odometry, radar RIO correspondence/uncertainty, Doppler radar-LiDAR bridge SLAM, raw GNSS factor fusion, wheel/vehicle-motion factors, radar-GNSS/visual mapping, fusion SLAM, robust backends, collaborative SLAM, alternative sensors, lifelong localization, static-map lifecycle/removal, dynamic-aware heterogeneous LiDAR map merging, LAMM multi-session point-cloud map merging, MapEval point-cloud map-quality evaluation, map cleaning, SLAM Toolbox, NDT variants, current benchmark pages, and neural/Gaussian SLAM taxonomy |
| ML-related SLAM research scope | 30-autonomy-stack/localization-mapping/overview/ml-related-slam-research-scope.md | Cross-section architecture for learned registration, learned loop retrieval, semantic/dynamic SLAM, neural implicit/Gaussian maps, dynamic residual removal, layered removal labels, motion/permanence/map-eligibility separation, static-but-transient point quarantine, multi-session map merging, learned map priors, and map-scale semantic segmentation handoffs |
| GLIM/GTSAM pipeline hub | 30-autonomy-stack/localization-mapping/slam-methods/glim-gtsam-pipeline-hub.md | Cross-section route linking GLIM stages to GTSAM factor graph objects, Bayes trees, Hessians, sparse elimination, marginalization, robust losses, and diagnostic KB pages |
| SLAM coverage audit | 30-autonomy-stack/localization-mapping/slam-methods/coverage-audit-2026.md | Source-backed backlog plus May 2026 discovery sweeps: LVI-SAM, FAST-LIVO/R3LIVE, KISS-SLAM, MOLA, robust/certifiable PGO, C-SLAM systems, degeneracy-robust LIO, event/thermal/UWB VIO and localization, raw GNSS factor fusion, wheel/vehicle-motion factors, radar-to-LiDAR map matching, 4D radar, Gaussian/foundation SLAM, and current benchmarks |
| AV / indoor / outdoor selection | 30-autonomy-stack/localization-mapping/slam-methods/av-indoor-outdoor-decision-matrix.md | Method fit by GNSS availability, dynamics, map dependence, compute budget, and safety criticality |
| Benchmarks and datasets | 30-autonomy-stack/localization-mapping/slam-methods/benchmarking-metrics-datasets.md | ATE/RPE, KITTI drift, loop closure, MapEval-style point-cloud map quality, dynamic-scene metrics, COSMO-Bench, LaMAria, Hilti x Trimble 2026, ScaleMaster, Oxford Spires, IILABS 3D, SMapper-light, FusionPortableV2, S3E, KITTI/KITTI-360, EuRoC, TUM, Oxford, Boreas, MulRan |
| Open-source stacks | 30-autonomy-stack/localization-mapping/slam-methods/open-source-stack-comparison.md | ORB-SLAM3, RTAB-Map, Cartographer, SLAM Toolbox, OpenVINS, Kimera, KISS-ICP, LIO-SAM, FAST-LIO2, GLIM, GTSAM, Open3D |
| Practical ROS mapping and localization | 30-autonomy-stack/localization-mapping/slam-methods/slam-toolbox.md | ndt-variants-and-ndt-maps.md, ndt.md, open-source-stack-comparison.md, and av-indoor-outdoor-decision-matrix.md |
| Robust and collaborative SLAM backends | 30-autonomy-stack/localization-mapping/slam-methods/robust-pgo-gnc-risam.md, 30-autonomy-stack/localization-mapping/slam-methods/certifiable-pose-graph-optimization.md, 30-autonomy-stack/localization-mapping/slam-methods/kimera-rpgo-pcm.md, 30-autonomy-stack/localization-mapping/slam-methods/distributed-multi-robot-pgo.md, 30-autonomy-stack/localization-mapping/slam-methods/kimera-multi.md, 30-autonomy-stack/localization-mapping/slam-methods/covins-covins-g.md, 30-autonomy-stack/localization-mapping/slam-methods/d2slam.md | GNC/Black-Rangarajan/riSAM, SE-Sync/Shonan-style certifiable PGO, pairwise consistency loop verification, distributed PGO, and full collaborative SLAM systems |
| Classical SLAM foundations | 30-autonomy-stack/localization-mapping/slam-methods/graphslam-pose-graph-optimization.md | ekf-slam.md, fastslam-particle-slam.md, bundle-adjustment-slam.md, factor-graph-isam2-gtsam.md, lidar-bundle-adjustment-factors.md, scan-context-family.md, loop-closure-place-recognition.md, occupancy-grid-tsdf-esdf-mapping.md |
| Point-cloud registration | 30-autonomy-stack/localization-mapping/slam-methods/gicp-vgicp.md | icp.md, point-to-plane-icp.md, ndt.md, ndt-variants-and-ndt-maps.md, continuous-time-registration.md |
| 3D LiDAR SLAM | 30-autonomy-stack/localization-mapping/slam-methods/kiss-icp.md | loam.md, lego-loam.md, hdl-graph-slam.md, ct-icp.md, lio-sam.md, fast-lio-fast-lio2.md, point-lio.md, glim.md, cartographer-3d.md, suma.md |
| Visual and visual-inertial SLAM | 30-autonomy-stack/localization-mapping/slam-methods/orb-slam2-orb-slam3.md | lsd-slam-dso.md, svo.md, vins-mono-vins-fusion.md, openvins.md, okvis2-x.md, kimera-vio.md, event-camera-vio-slam.md, thermal-inertial-slam.md, droid-slam.md, dpvo.md, mast3r-slam.md, lamaria-city-scale-visual-inertial-slam-benchmark.md, hilti-trimble-slam-challenge-2026.md, scalemaster-benchmark.md |
| Indoor and dense SLAM | 30-autonomy-stack/localization-mapping/slam-methods/rtab-map.md | kinectfusion.md, elasticfusion.md, bundlefusion.md, imap.md, nice-slam.md, co-slam-eslam.md, nerf-slam.md |
| Learned, semantic, and Gaussian SLAM | 30-autonomy-stack/localization-mapping/slam-methods/splatam.md | lo-net-learned-lidar-odometry.md, regformer-learned-registration.md, semantic-slam.md, dynamic-object-aware-slam.md, object-level-slam.md, multi-agent-neural-gaussian-slam.md, gs-slam-monogs.md, photo-slam.md, neural-gaussian-slam-surveys.md |
| Outdoor Gaussian, radar, and degraded-sensor SLAM | 30-autonomy-stack/localization-mapping/slam-methods/splat-loam.md | gigaslam.md, wildgs-slam.md, splat-slam.md, s3po-gs.md, hi-slam2.md, segs-slam.md, gaussian-lic.md, rmgs-slam.md, gs-livm.md, vigs-slam.md, dynamic-4d-gaussian-slam.md, radarsplat-rio.md, 4d-imaging-radar-rio-slam.md, radar-rio-correspondence-uncertainty.md, doppler-radar-lidar-slam.md, gvins-glio-gnss-raw-factor-fusion.md, wheel-odometry-vehicle-motion-factors.md, radar-to-lidar-map-localization.md, radar-odometry-radar-slam.md, radar-place-recognition-4dral-sherloc.md, radar-inertial-odometry.md, radar-lidar-inertial-fusion.md, uwb-radio-ranging-slam.md, mm-lins.md |
| Dynamic map cleaning | 30-autonomy-stack/localization-mapping/slam-methods/lidar-map-cleaning-dynamic-removal.md | erasor.md, removert.md, mapcleaner.md, erasor-plus-plus.md, 4dndf.md, freedom-dynamic-object-removal.md, beautymap.md, raymoval.md, static-lio-dynamic-points-removal.md, moves-and-label-free-map-cleaning.md, potentially-dynamic-object-removal-ground-projection.md, uni-mapper-dynamic-aware-lidar-map-merging.md, lamm-multi-session-point-cloud-map-merging.md, benchmark coverage, dynamic residual removal, static-but-transient quarantine, reason-coded release decisions, static-but-wrong map exclusion, static-map preservation, heterogeneous/multi-session map merging, and dynamic-object removal risks |
| Lifelong and alternative localization | 30-autonomy-stack/localization-mapping/slam-methods/lt-mapper-khronos-lifelong-mapping.md, 30-autonomy-stack/localization-mapping/slam-methods/lifelong-3d-map-version-control.md, 30-autonomy-stack/localization-mapping/slam-methods/uni-mapper-dynamic-aware-lidar-map-merging.md, 30-autonomy-stack/localization-mapping/slam-methods/lamm-multi-session-point-cloud-map-merging.md, 30-autonomy-stack/localization-mapping/slam-methods/rtmap-dufomap-recursive-maintenance.md, 30-autonomy-stack/localization-mapping/slam-methods/gpr-localization-ground-encoding.md, 30-autonomy-stack/localization-mapping/slam-methods/radar-teach-repeat-localization.md | Long-term metric-semantic mapping, reconstructable map version control, dynamic-aware heterogeneous-LiDAR map merging, large-scale multi-session point-cloud map merging, recursive map maintenance, ground-penetrating radar localization, and radar teach-and-repeat fallbacks for adverse weather or changed scenes |
Localization & mapping
| Topic | Primary | Supporting |
|---|---|---|
| Mapping & localization overview | 30-autonomy-stack/localization-mapping/overview/mapping-and-localization.md | MapTR, NMP, Tesla/Mobileye, SLAM |
| ML-related SLAM research scope | 30-autonomy-stack/localization-mapping/overview/ml-related-slam-research-scope.md | Learned SLAM research scope for semantic aggregated maps: learned registration, learned place recognition, semantic/dynamic SLAM, neural implicit and Gaussian maps, point-cloud removal, motion/permanence/map-eligibility separation, static-but-transient quarantine, map priors, and segmentation handoff |
| Infrastructure-aided localization | 30-autonomy-stack/localization-mapping/overview/infrastructure-aided-localization.md | UWB anchors, fiducials, RFID/BLE, Wi-Fi RTT, magnetic maps, surveyed reflectors, 5G NR/mmWave positioning, measurement contracts, lifecycle controls, and fiducial/corner pose evidence routing |
| Map-free driving for airports | 30-autonomy-stack/localization-mapping/maps/map-free-driving.md | Three-layer map, AIXM prior, 10-25x faster deployment |
| HD map standards (airside) | 30-autonomy-stack/localization-mapping/maps/hd-map-standards-airside.md | OpenDRIVE, AMDB/AMXM, NDS, NOTAM integration, AIRAC cycle |
| Neural online mapping SOTA | 30-autonomy-stack/localization-mapping/maps/neural-online-mapping-sota.md | MapTracker, StreamMapNet, NMP, topology (TopoMLP, LaneSegNet) |
| LiDAR SLAM algorithms | 30-autonomy-stack/localization-mapping/overview/lidar-slam-algorithms.md | KISS-ICP, LIO-SAM, FAST-LIO2, Point-LIO, degeneracy handling |
| Semantic mapping & learned priors | 30-autonomy-stack/localization-mapping/maps/semantic-mapping-learned-priors.md | Neural Map Prior (NMP +5.4 mAP), PriorDrive, T2SG topology graphs, conformal map uncertainty, fleet-based incremental updates, 7-layer semantic map, learned-prior acceptance rules, multi-airport LoRA adapters |
| HD map change detection & maintenance | 30-autonomy-stack/localization-mapping/maps/hd-map-change-detection-maintenance.md | Point cloud differencing, semantic change detection, RTMap (ICCV 2025 centimeter-level), Bayesian fleet consensus, AIRAC integration, temporal decay models, light-map alternative (~720 KB), NMP implicit maintenance, OTA canary deployment, construction zone detection, cost 60-80% reduction vs manual re-survey, $45-70K/28 weeks |
| Moved-object and map-change datasets | 30-autonomy-stack/localization-mapping/maps/moved-object-and-map-change-datasets.md | RTMap/ExelMap-style change detection, 3RScan/Objects Can Move, TbV, POCD, FOD-A, dynamic-map benchmarks, and fleet-consensus validation |
| LiDAR place recognition & re-localization | 30-autonomy-stack/localization-mapping/overview/lidar-place-recognition-relocalization.md | Scan Context (<5ms CPU) + MinkLoc3D (97.5% recall@1, 15ms GPU) two-stage pipeline, PointNetVLAD, LoGG3D-Net, LCDNet (integrated pose), PPT few-shot, BEVPlace, FAISS million-scale retrieval (<1ms), GTSAM loop closure factors, kidnapped robot recovery, fleet shared descriptors, identical-stands disambiguation, seasonal databases, $33-57K/12-16 weeks |
| Robust state estimation & multi-sensor fusion | 30-autonomy-stack/localization-mapping/overview/robust-state-estimation-multi-sensor.md | ESKF (Error-State Kalman Filter) with quaternion error parameterization, chi-squared innovation gating, Mahalanobis sensor validation, multi-hypothesis tracking (IMM), GPS-denied dead-reckoning budgets, adaptive noise estimation (Sage-Husa), covariance management, fleet-level state consistency, <0.5ms per update on Orin, robot_localization integration |
| Real-time occupancy grid mapping | 30-autonomy-stack/localization-mapping/maps/realtime-occupancy-grid-mapping.md | Log-odds Bayesian update, OctoMap/VDBFusion/nvblox comparison, GPU raycasting (CUDA), multi-LiDAR fusion (4-8 sensors at 10Hz), dynamic object separation, multi-resolution grids (0.1-0.8m), TSDF/ESDF for planning, costmap generation for Frenet planner, fleet-shared occupancy over 5G, airside-specific (aircraft stands, jet blast), $25-40K |
| HD map construction pipeline | 30-autonomy-stack/localization-mapping/maps/map-construction-pipeline.md | End-to-end offline map building: survey drive planning (3 drive patterns), multi-session SLAM (FAST-LIO2+GTSAM), point cloud post-processing (dynamic object removal via multi-session voting), geodetic alignment (RTK+GCPs ±5-10cm global), OpenLiDARMap map-prior georeferencing for GCP-sparse sites, FlexCloud GNSS/reference-trajectory drift correction, source-map acceptance package before semantic segmentation, AMDB overlay and co-registration, automated annotation (SAM+CLIP 85-92% accuracy), Lanelet2 generation, QA validation (20 automated checks), map packaging and OTA deployment, DVC version control, map CI/CD pipeline, 5-7 days per airport at $20-35K, scaling to $12-19K/airport at 20 airports |
| Potentially dynamic object map policy | 30-autonomy-stack/localization-mapping/maps/potentially-dynamic-object-map-policy.md | Airside object-layer policy separating permanent static map content, movable-static/current occupancy, temporary overlays, FOD/hazard handling, artifacts, and the permanence evidence ladder for stationary-but-transient objects |
| Production LiDAR-to-map localization | 30-autonomy-stack/localization-mapping/overview/production-lidar-map-localization.md | Runtime scan-to-map matching pipeline: ICP/GICP/VGICP/NDT algorithm comparison, multi-resolution coarse-to-fine (NDT→VGICP), eigenvalue-based degeneracy detection and handling, airside-specific challenges (40-70% dynamic content at stands, jet blast shimmer, ground reflectivity), multi-LiDAR fusion strategies (merge-then-match, match-then-fuse, selective), GTSAM factor graph integration with adaptive noise models, 5-level fallback hierarchy (VGICP→NDT→GPS→dead reckoning→safe stop), learned registration (GeoTransformer for cold start), Orin GPU deployment (15-25ms typical), $30-53K/12 weeks |
| Map tile versioning & distribution | 30-autonomy-stack/localization-mapping/maps/map-tile-versioning-distribution.md | Map lifecycle from build to vehicle: spatial tiling (50-200m tiles), content-addressable versioning (SHA-256 Merkle tree), differential updates (bsdiff, 2-8% of full tile), distribution over airport 5G (<30s/tile), NVMe vehicle-side storage, atomic map swap protocol (zero perception gaps), AIRAC 28-day cycle integration, cryptographic signing (Ed25519), fleet version synchronization, in-flight consistency, <500 MB/month fleet updates, $40-65K/14 weeks |
AV platform
| Topic | Primary | Supporting |
|---|---|---|
| NVIDIA Orin | 20-av-platform/compute/nvidia-orin-technical.md | 275 TOPS, 8 power modes, benchmarks |
| NVIDIA Thor | 20-av-platform/compute/nvidia-drive-thor.md | ~1000 TOPS, FP8, OEM commitments |
| TensorRT deployment | 20-av-platform/compute/tensorrt-deployment-guide.md | DLA, quantization, Lidar_AI_Solution |
| Hesai LiDAR | 20-av-platform/sensors/hesai-lidar.md | XT32, AT128 ASIL-B, FMC500 SoC |
| RoboSense LiDAR | 20-av-platform/sensors/robosense-lidar.md | RSHELIOS, RSBP, 7-sensor layout |
| 4D radar | 20-av-platform/sensors/4d-radar.md | Continental ARS548, weather immunity |
| Visible cameras | 20-av-platform/sensors/visible-cameras.md | Global vs rolling shutter, HDR/LFM, lens/FOV, trigger/PTP, ISP/RAW, cleaning, heating, and weather integration |
| IMU, GNSS, and RTK hardware | 20-av-platform/sensors/imu-gnss-rtk.md | Receiver/IMU classes, PPS/PTP wiring, antenna lever arms, correction transport, outage modes, spoofing/jamming health |
| Thermal/IR cameras | 20-av-platform/sensors/thermal-ir-cameras.md | FLIR Boson 640, LWIR fusion, night personnel, jet blast |
| Calibration bay fixtures | 20-av-platform/sensors/calibration-bay-fixtures.md | Physical calibration-bay design: surveyed bay frames, turntable/vehicle datum, target inventory, LWIR/radar/LiDAR/camera fixtures, capture manifests, residual evidence, release gates, and airside transfer notes |
| Multi-LiDAR calibration | 20-av-platform/sensors/multi-lidar-calibration.md | Target-based + targetless (ICP, feature, learning-based), GTSAM-integrated online refinement, thermal drift compensation (-10C to +50C), PTP/PPS synchronization, overlap optimization for 4-8 RoboSense, calibration health monitoring, ISO 3691-4 traceability, 400-800h/year labor savings for 20+ vehicle fleet |
| Sensor-to-algorithm readiness | 20-av-platform/sensors/sensor-to-algorithm-readiness-contract.md | Pre-algorithm contract for calibration, timestamp, TF, preprocessing, health, provenance, and reject/degrade gates before perception, fusion, SLAM, tracking, occupancy, mapping, and planning consume sensor data |
| Sensor degradation & health monitoring | 20-av-platform/sensors/sensor-degradation-health-monitoring.md | Degradation taxonomy (optical/mechanical/environmental/electronic), 10 airside contamination sources, per-sensor diagnostics (LiDAR 7-check, radar SNR/coverage, thermal NUC/dead pixel, camera exposure/blur), cross-sensor consistency scoring, EMA-based temporal tracking with z-score anomaly, response matrix (4 sensors × 4 severity), fleet health analytics (zone correlation, seasonal patterns), predictive maintenance (linear extrapolation), cleaning schedules, 1 Hz ROS monitoring at <2ms, $35K/11 weeks |
| Automated sensor cleaning | 20-av-platform/sensors/automated-sensor-cleaning.md | Physical self-maintenance for 16-20 hr/day tarmac ops: cleaning modality comparison (air curtains, air burst, wipers, washer fluid, ultrasonic, heated windows, hydrophobic coatings, UV photocatalytic), contamination-to-cleaning mapping (de-icing glycol requires chemical cleaning — air jets spread it), per-sensor architecture (germanium thermal windows air-only, no wipers), health monitor closed-loop integration, power/weight budget (15-40W, 1.5-3.0 kg), $200-500/vehicle hardware, 15-25% availability improvement, 60-80% fewer depot cleaning visits |
| Solid-state LiDAR & photonics | 20-av-platform/sensors/solid-state-lidar-photonics.md | FMCW vs ToF measurement principles, silicon photonics integration (SiPh LiDAR-on-chip), OPA beam steering (GHz point-to-point, 0.01-0.05° angular resolution), MEMS mirror reliability, flash LiDAR for docking, per-point velocity (jet blast detection, zero-latency approaching-object detection), 1550nm eye safety (100x margin), Aeva Atlas/Voyant Helium/SiLC comparison, 50-200x longer MTBF (100K+ hrs), $150-450K/year fleet savings, Orin GTSAM velocity factor, adaptive resolution for active perception, phased migration strategy, $110-175K over 48 weeks |
| LiDAR ghost and multipath artifacts | 20-av-platform/sensors/lidar-ghost-multipath-artifacts.md | Wet surfaces, aircraft skins, glass, retroreflector bloom, sun/receiver saturation, multi-return ambiguity, and cross-sensor checks |
| Energy-efficient inference 24/7 | 20-av-platform/compute/energy-efficient-inference-24-7.md | Orin 15W/30W/50W power modes vs throughput, dynamic model switching (40-60% time in low-complexity), thermal throttling curves (-10C to +50C tarmac), battery-aware compute (SoC-correlated power budgets), DLA+GPU concurrent scheduling, per-model watt measurements, sleep/wake with <500ms wake-up, fleet-level energy optimization, 8-15% more daily operating hours, 12-18C lower junction temp |
| Edge-cloud hybrid inference | 20-av-platform/compute/edge-cloud-hybrid-inference.md | Three-tier architecture (on-vehicle Orin + airport MEC edge + cloud), model placement decision framework, split inference patterns, bandwidth/latency analysis, NVIDIA Triton on edge servers, graceful degradation (vehicle always autonomous), security, cost-benefit ($2,500/vehicle for shared edge vs $2,000-5,000 per Thor upgrade), industry approaches, airport advantage (bounded geography + private 5G) |
| Airport 5G | 20-av-platform/networking-connectivity/airport-5g-cbrs.md | 20-av-platform/networking-connectivity/airport-5g-case-studies.md |
| Deterministic networking (TSN) | 20-av-platform/networking-connectivity/deterministic-networking-tsn.md | IEEE 802.1 TSN standards (gPTP <100ns sync, TAS time-aware scheduling, FRER redundancy, frame preemption), mixed-criticality traffic classes (safety <100μs, sensors <5ms, best-effort), CAN bus migration (50-200x latency improvement for safety messages), zonal architecture, automotive TSN silicon (NXP SJA1110, Marvell 88Q6113), CAN-TSN gateway (NXP S32G3), 5G TSN bridge for V2X, Orin native TSN support, ASIL decomposition via TSN isolation, $230-440/vehicle hardware, $53-87K implementation |
Safety & certification
| Topic | Primary | Supporting |
|---|---|---|
| ISO 3691-4 | 60-safety-validation/standards-certification/iso-3691-4-deep-dive.md | 27 functions, $130K-380K |
| Full certification guide | 60-safety-validation/standards-certification/certification-guide.md | UL 4600, AMLAS, ISO 26262 |
| Regulatory trajectory | 80-industry-intel/regulations/regulatory-trajectory-deep-dive.md | FAA, EASA, CAAS, predicted timeline |
| Safety incidents | 60-safety-validation/safety-case/safety-incidents-lessons.md | Cruise, Waymo, Tesla, Uber ATG |
| Failure modes | 60-safety-validation/safety-case/failure-modes-analysis.md | SOTIF, hallucination taxonomy |
| Simplex architecture | 60-safety-validation/runtime-assurance/simplex-safety-architecture.md | RSS, OOD detection, ROS dual-stack |
| Ground crew safety | 70-operations-domains/airside/safety/ground-crew-pedestrian-safety.md | 27K accidents/yr, hi-vis paradox |
| Insurance & liability | 80-industry-intel/regulations/insurance-liability-airside.md | EU PLD, $35M exposure |
| Functional safety software | 60-safety-validation/standards-certification/functional-safety-software.md | MISRA C, ISO 26262 Part 6, static analysis, CI/CD, ROS safety patterns |
| Scenario taxonomy & edge cases | 60-safety-validation/verification-validation/airside-scenario-taxonomy.md | ISO 34502 adapted for airside, SOTIF hazard catalog (H1-H8+), 115 functional scenarios, ODD definition, Pegasus 6-layer, STPA, risk matrix, regulatory mapping |
| Testing & validation methodology | 60-safety-validation/verification-validation/testing-validation-methodology.md | V-model, scenario-based testing (ASAM OpenSCENARIO 2.0), coverage metrics (N-wise covering arrays), corner case/adversarial testing (CMA-ES falsification, LLM scenario generation, metamorphic testing), SIL/HIL/VIL, statistical safety (Zhao-Weng, Bayesian), shadow mode, regression/CI/CD, digital twin, airside test protocols, $105K first airport |
| LiDAR artifact removal validation | 60-safety-validation/verification-validation/robustness/lidar-artifact-removal-validation.md | Raw-vs-filtered evidence, do-not-delete hazard tests, weather/ghost/dynamic-object labels, localization observability, ODD degradation, and SOTIF argumentation |
| Airside dynamic map-cleaning benchmark | 60-safety-validation/verification-validation/airside-dynamic-map-cleaning-benchmark.md | False-deletion, false-retention, moved-object, FOD, construction, equipment, and localization-regression tests for map cleaning |
| Map publication gates for dynamic removal | 60-safety-validation/verification-validation/map-publication-gates-dynamic-object-removal.md | Release gates for cleaned map publication: static preservation, dynamic rejection, FOD retention, localization replay, manifest contracts, and evidence conflict checks across motion, semantics, persistence, operations, and reviewer policy |
| Runtime verification & monitoring | 60-safety-validation/runtime-assurance/runtime-verification-monitoring.md | STL monitors (<1ms, 20 airside specs), OOD detection (energy+Mahalanobis+ensemble, 95-98% AUROC), maximally permissive shields (1-5% intervention), safety MCU (STM32H725), METAR ODD monitoring, WCET <5.5ms, ISO 26262 ASIL decomposition, UL 4600 compliance, DO-178C credit, fleet anomaly correlation, $115-200K/32 weeks |
| Online perception monitoring & ODD enforcement | 60-safety-validation/runtime-assurance/online-perception-monitoring-odd-enforcement.md | ML-specific silent degradation detection, input/output drift, cross-modal consistency, OOD integration, ODD boundary state machine, Perception Health Score, calibration drift, temporal anomaly detection, MLOps-scale runtime monitoring handoff, <5ms total on Orin |
| Formal verification of neural networks | 60-safety-validation/verification-validation/formal-verification-neural-networks.md | SMT/MILP complete verification (<100K params), alpha-beta-CROWN over-approximation (millions of params, VNN-COMP winner), IBP/SABR certified training, Lipschitz bounds for safety margins, layered strategy: complete for safety-critical (policy, CBF, Simplex), scalable for perception (PointPillars, CenterPoint), runtime for residual, ISO 3691-4/UL 4600/EU AI Act compliance |
| Fail-operational architecture | 60-safety-validation/runtime-assurance/fail-operational-architecture.md | 1oo2D, TMR, monitor-actuator patterns, dual-Orin compute, Orin FSI (DCLS R52), ASIL decomposition, sensor/actuator/power/CAN redundancy, degradation tiers, MRC planning, airside-specific (runway incursion HW geofence, jet blast hardening, EMI), $155-260K phased implementation |
| Weather-adaptive ODD management | 60-safety-validation/runtime-assurance/weather-adaptive-odd-management.md | 5-level ODD (A-E) with asymmetric transitions (fast degradation, slow recovery), METAR/TAF/ATIS automated parsing, on-vehicle environmental sensing (LiDAR return rate→visibility), fleet consensus, capability curves (sensor performance vs weather), continuous speed envelope, jet blast zone integration (ADS-B+thermal), seasonal adaptation profiles, dawn/dusk transition management, ISO 34502/21448/3691-4 compliance, EU AI Act transparency, $30-50K/8-12 weeks |
Planning, VLA & scene understanding
| Topic | Primary | Supporting |
|---|---|---|
| VLA for driving | 30-autonomy-stack/vla-vlm/vla-for-driving.md | Alpamayo, RT-2, PaLM-E, teacher-student distillation |
| Alpamayo setup | 30-autonomy-stack/vla-vlm/alpamayo-setup.md | Camera-only, non-commercial, 10B params |
| VLM scene understanding | 30-autonomy-stack/vla-vlm/vlm-scene-understanding.md | DriveVLM, DriveLM, NOTAM interpretation, turnaround assessment, FOD classification, VLM as 1-2Hz co-pilot |
| Spatial foundation models | 30-autonomy-stack/vla-vlm/spatial-foundation-models-airport.md | 4M unified multimodal, SpatialVLM spatial reasoning, RT-2/RT-X robotics transformers, Octo open-source policy, pi0 flow matching, HPT cross-embodiment, precision docking with spatial VLMs, gate identification, FOD detection/characterization, two-tier deployment (cloud+edge), distillation for Orin, in-context learning for new airports, Simplex integration, $55-95K phased |
| Neural motion planning | 30-autonomy-stack/planning/neural-motion-planning.md | SparseDrive/SparseDriveV2, DiffusionDrive/DiffusionDriveV2, GameFormer, NAVSIM PDMS/EPDMS, Bench2Drive caveats, Simplex safety integration |
| Frenet augmentation | 30-autonomy-stack/planning/frenet-planner-augmentation.md | Augmenting a classical Frenet planner |
| Motion prediction | 30-autonomy-stack/planning/motion-prediction.md | Trajectory prediction, interaction modeling |
| LLM reasoning for planning | 30-autonomy-stack/planning/llm-reasoning-planning.md | Chain-of-thought, interpretable decisions |
| Diffusion trajectory planning | 30-autonomy-stack/planning/diffusion-trajectory-planning.md | Diffuser, DiffusionDrive, DiffusionDriveV2, truncated diffusion, anchor-based planning, RL-constrained trajectory selection |
| Safety-critical planning (CBF) | 30-autonomy-stack/planning/safety-critical-planning-cbf.md | Control Barrier Functions, CBF-QP safety filter, neural CBF synthesis, game-theoretic interaction (GameFormer, GIME), multi-agent CBFs (GCBF+), HJ reachability, CBF-Simplex integration, airside safety formulations |
| Neuro-symbolic scene graphs | 30-autonomy-stack/planning/neuro-symbolic-scene-graphs.md | Driving scene graphs, GNN interaction (LaneGCN, HiVT, HDGT), knowledge graphs for traffic rules, STL-constrained planning, compositional reasoning, LLM-symbolic hybrid, airside right-of-way encoding, NOTAM rule injection, interpretable decisions |
| Causal reasoning & counterfactuals | 30-autonomy-stack/planning/causal-reasoning-counterfactual.md | SCMs for driving, Pearl's 3 levels, counterfactual trajectory analysis, Halpern-Pearl causation, NOTEARS causal discovery, IRM cross-airport transfer, off-policy evaluation, LLM+SCM hybrid, KING counterfactual generation, EU PLD 2024/2853 compliance, causal ROS node at 2 Hz, $40-65K Phase 1+2 |
| RL driving policy | 30-autonomy-stack/planning/reinforcement-learning-driving-policy.md | CaRL (CoRL 2025 SOTA, PPO + simple rewards), IQL (best offline RL), SAC/TD3/TQC/CrossQ, BC→offline RL→online RL pipeline, safe RL (CPO, Lagrangian, CBF filter), privileged-to-sensor distillation, policy head <0.5ms Orin, $45-75K/32 weeks |
| Imitation learning & behavioral cloning | 30-autonomy-stack/planning/imitation-learning-behavioral-cloning.md | BC from teleop, MDN multimodal BC, Diffusion BC (DDIM 3-5 steps), DAgger with Frenet expert, MaxEnt IRL cost learning, GAIL, style-conditioned multi-operator BC, CBF safety filtering, Simplex integration, $35-55K/10-14 weeks |
| Joint prediction-planning | 30-autonomy-stack/planning/joint-prediction-planning.md | Predict-then-plan failure modes, PDM-Closed baseline, conditional prediction, game-theoretic (Stackelberg, level-K), contingency planning, occupancy flow scoring, NAVSIM/nuPlan benchmarks, Frenet planner augmentation with prediction costs, airside interaction modeling, 50-100ms on Orin |
| Autonomous docking & precision positioning | 30-autonomy-stack/planning/autonomous-docking-precision-positioning.md | Two-phase architecture (coarse Frenet → fine docking), visual servoing (IBVS/PBVS), LiDAR ICP template alignment (+-1-2cm), AprilTag fiducials (+-0.5cm at 2m), MPC docking controller (CasADi 2-5ms), impedance control for pushback contact, per-GSE tolerances (+-5cm belt loader to +-30cm fuel truck), third-generation tug crab steering advantage, safety PLC + personnel exclusion zones, 20 key takeaways, $53-90K/12-18 weeks |
Airport operations
| Topic | Primary | Supporting |
|---|---|---|
| Industry overview | 70-operations-domains/airside/operations/industry-overview.md | All competitors, regulatory gaps |
| Airport data APIs | 70-operations-domains/airside/operations/airport-data-integration.md | 70-operations-domains/airside/operations/airport-data-systems-detailed.md (real endpoints) |
| FOD & jet blast | 70-operations-domains/airside/operations/fod-and-jetblast.md | B737 148m zone, CFD tables |
| Turnaround prediction | 70-operations-domains/airside/operations/turnaround-prediction.md | Moonware HALO, Assaia |
| Pushback systems | 70-operations-domains/airside/operations/pushback-systems.md | Mototok, TaxiBot, WheelTug |
| Electric GSE market | 70-operations-domains/airside/operations/electric-gse-market.md | $2.8B→$5.2B, autonomy rankings |
| Aviation ecosystem | 70-operations-domains/airside/operations/aviation-ground-ops-ecosystem.md | Strategic context, business case |
| Battery & charging | 70-operations-domains/airside/operations/battery-charging-infrastructure.md | LiFePO4, 0.84yr payback, autonomous self-charging |
| Ground control instructions | 70-operations-domains/airside/operations/ground-control-instructions.md | A-CDM/A-SMGCS integration, D-TAXI, NOTAM parsing, marshaller gesture recognition, NLU, instruction-to-trajectory |
Deployment & operations
| Topic | Primary | Supporting |
|---|---|---|
| Deployment playbook | 70-operations-domains/deployment-playbooks/deployment-playbook.md | 4,500 lines, full checklists |
| Shadow mode | 60-safety-validation/verification-validation/shadow-mode.md | Tesla/Waymo/comma approaches |
| OTA & fleet management | 50-cloud-fleet/ota/ota-fleet-management.md | Canary deployment, A/B testing |
| Production ML | 40-runtime-systems/ml-deployment/production-ml-deployment.md | TensorRT, Triton, GPU reliability, runtime promotion evidence, monitoring by MLOps scale, and drift evidence contracts for release tickets, ODD-cell quarantine, safety-case updates, and platform SLOs |
| Fleet dispatch | 50-cloud-fleet/fleet-management/fleet-management-dispatch.md | VRPTW, A-CDM triggers |
| Multi-airport adaptation | 70-operations-domains/deployment-playbooks/multi-airport-adaptation.md | AMDB bootstrapping, PointLoRA fine-tuning (500 labels), GNSS multipath mapping, 8-week onboarding, $75-150K per airport |
| HMI & operator interface | 40-runtime-systems/monitoring-observability/hmi-operator-interface.md | Dashboard design, trust calibration, 4-mode control, handoff procedures, operator training, incident reporting |
| Teleoperation | 40-runtime-systems/monitoring-observability/teleoperation-systems.md | Fernride, Waymo 1:41 ratio |
| Workforce transition | 70-operations-domains/deployment-playbooks/workforce-transition.md | 1.5-2M workers affected, union considerations, retraining |
| CI/CD & DevOps pipeline | 40-runtime-systems/ml-deployment/av-cicd-devops-pipeline.md | End-to-end AV CI/CD: code CI, ML model CI, SIL/HIL/VIL gates, map/config CI, artifact versioning, fleet deployment, ML regression detection, safety assurance, and S0-S5 MLOps pipeline architecture |
| Fleet TCO & business case | 70-operations-domains/airside/business-case/fleet-tco-business-case.md | Per-vehicle CAPEX ($95-210K floor), OPEX breakdown, 3-shift labor savings ($150K/year), NPV $45-80M at 200 vehicles, break-even Year 2-4, RaaS $10-14K/month, certification cost $530K-1.95M across 5 jurisdictions, UISEE 40-60% cost advantage threat, airport cluster strategy |
| EV fleet energy co-optimization | 50-cloud-fleet/fleet-management/ev-fleet-energy-co-optimization.md | Joint charging-routing-task EVRP optimization, LiFePO4 degradation models (cycle counting, throughput, temperature), optimal C-rate selection, V2G for airports (~1-2 MWh dispatchable storage, $50-200/MWh demand response), grid-aware scheduling (demand charge management), stochastic EVRP under uncertainty, MILP/RL/MPC approaches, OCPP 2.0.1 integration |
| Fleet anomaly root-cause attribution | 50-cloud-fleet/observability/fleet-anomaly-root-cause-attribution.md | Automated causal attribution for fleet-level anomalies: CUSUM/EWMA monitoring, hierarchical anomaly detection, MLOps observability by scale, causal discovery, Shapley attribution, Bayesian diagnosis, OTA regression, map staleness, environmental correlation, MTTR reduction |
| Fleet predictive maintenance | 50-cloud-fleet/fleet-management/fleet-predictive-maintenance.md | PHM framework (ISO 13381, 4-level architecture), Weibull failure analysis (LiDAR β=1.8-2.2/25-40K hrs, motors β=3.5/40-60K hrs), correlated failure modes (de-icing, salt spray, heat events), ML prediction (LSTM/XGBoost/autoencoder anomaly), multi-echelon spare parts inventory (4-level), cold-start sizing for new airports, joint maintenance-operations scheduling (CP-SAT), fleet availability modeling (95%+ vehicle, 98%+ fleet), seasonal profiles, ROS diagnostics integration, $7-19.5K/vehicle/year maintenance cost, 30-40% cost reduction with predictive vs reactive, $50-80K implementation |
Foundation overview entry points
| Topic | Primary | Scope |
|---|---|---|
| Probability and statistics foundations | 10-knowledge-base/probability-statistics/overview.md | Uncertainty, likelihoods, covariance, gates, robust statistics, calibration, and decision thresholds |
| Optimization foundations | 10-knowledge-base/optimization/overview.md, Nonlinear Solver Diagnostics Crosswalk | Residual objectives, Jacobians, manifold linearization, globalization, solver patterns, and failure triage across residual, scaling, damping, rank, covariance, and backend causes |
| Numerical linear algebra foundations | 10-knowledge-base/numerical-linear-algebra/overview.md, Sparse Estimation Backend Crosswalk, Nonlinear Solver Diagnostics Crosswalk | Factorization, conditioning, rank, sparsity, Schur complements, marginalization, covariance recovery, and sparse backend triage |
| State estimation foundations | 10-knowledge-base/state-estimation/overview.md | Filtering, smoothing, fusion, association, robust-loss covariance consistency, observability, integrity, and deployed estimator lifecycle |
| Geometry and sensor foundations | 10-knowledge-base/geometry-3d/overview.md | Frames, projection, fiducial/corner localization, two-view epipolar/homography verification, optical/scene-flow motion fields, Lie groups, registration, calibration, and sensor geometry |
| Mapping foundations | 10-knowledge-base/mapping/overview.md | Occupancy, semantic layers, volumetric maps, fusion policy, dynamic/static separation, and map QA |
| Sensor foundations | 10-knowledge-base/sensors/overview.md | Measurement likelihoods, thermal IR radiometry, ultrasonic proximity sensing, error budgets, observability limits, degradation modes, and modality handoff assumptions |
| Sensor readiness handoff | 20-av-platform/sensors/sensor-to-algorithm-readiness-contract.md | Operational bridge from sensor foundations into algorithm input acceptance gates |
| Signal processing foundations | 10-knowledge-base/signal-processing/overview.md | Sampling, filtering, FFT, radar processing, CFAR, aliasing, windowing, and clutter contracts |
| Controls foundations | 10-knowledge-base/controls/overview.md | Closed-loop tracking, vehicle dynamics, MPC/iLQR, constraints, actuator limits, and safety filters |
| Robotics foundations | 10-knowledge-base/robotics/overview.md | Robot/task vocabulary, route/behavior/motion-planning boundaries, Lanelet2 concepts, and embodiment assumptions |
| Systems engineering foundations | 10-knowledge-base/systems-engineering/overview.md | Timing, latency, validation metrics, release gates, observability, architecture contracts, and evidence flow |
| Machine learning foundations | 10-knowledge-base/machine-learning/overview.md | Learned representations, objectives, architectures, self-supervision, world models, AV data-evaluation contracts, evaluation, and deployment failure modes |
Mathematical foundations
| Topic | Primary |
|---|---|
| PointPillars | 10-knowledge-base/geometry-3d/pointpillars.md — tensor shapes, TensorRT |
| VQ-VAE / FSQ | 10-knowledge-base/machine-learning/vqvae-tokenization.md — straight-through estimator, codebook collapse |
| Transformers | 10-knowledge-base/machine-learning/transformer-world-models.md — causal attention, KV-cache, scaling laws |
| Diffusion models | 10-knowledge-base/machine-learning/diffusion-models.md — DDPM, DiT, flow matching |
| GTSAM | 10-knowledge-base/state-estimation/gtsam-factor-graphs.md — ISAM2, VGICP, neural factors |
| Probability and uncertainty | 10-knowledge-base/probability-statistics/gaussian-noise-covariance-information.md, 10-knowledge-base/probability-statistics/mahalanobis-chi-square-gating.md, 10-knowledge-base/probability-statistics/likelihood-map-mle-least-squares.md — Gaussian noise, covariance/information matrices, whitening, Mahalanobis gates, chi-square thresholds, NIS/NEES, likelihoods, MLE, MAP, and least-squares foundations |
| Robust statistics and multimodal beliefs | 10-knowledge-base/probability-statistics/robust-statistics-ransac-hypothesis-testing.md, 10-knowledge-base/probability-statistics/robust-losses-m-estimators-huber-cauchy-tukey-geman-mcclure.md, 10-knowledge-base/probability-statistics/mixture-models-multimodal-beliefs.md — RANSAC, hypothesis tests, Huber/Cauchy/Tukey/Geman-McClure M-estimators, Gaussian mixtures, mixture reduction, and multi-hypothesis localization/tracking |
| Estimator consistency and covariance integrity | 10-knowledge-base/state-estimation/robust-loss-covariance-consistency.md, 10-knowledge-base/state-estimation/slam-vio-observability-fej-nullspace-consistency.md, 10-knowledge-base/numerical-linear-algebra/square-root-information-and-covariance-recovery.md — covariance reporting after robust losses, gating, GNC, marginalization, weak modes, nullspaces, NIS/NEES, and release checks |
| Graphical models and information theory | 10-knowledge-base/probability-statistics/probabilistic-graphical-models-message-passing.md, 10-knowledge-base/probability-statistics/information-theory-for-perception-ml.md — factor graphs, Bayes nets, message passing, entropy, mutual information, KL divergence, active perception, and representation objectives |
| Calibration and uncertainty guarantees | 10-knowledge-base/probability-statistics/uncertainty-quantification-calibration-conformal.md — calibration error, reliability diagrams, conformal prediction, prediction sets, and safety-facing uncertainty contracts |
| ML evaluation and AV benchmark contracts | 10-knowledge-base/machine-learning/av-data-evaluation-fundamentals.md, 10-knowledge-base/machine-learning/evaluation-calibration-and-data-leakage-first-principles.md, 10-knowledge-base/systems-engineering/benchmarking-metrics-statistical-validity.md — split manifests, scenario/ODD coverage, leakage firewalls, calibration, statistical validity, and open-loop/pseudo/closed-loop evidence |
| Nonlinear optimization | Nonlinear Solver Diagnostics Crosswalk, 10-knowledge-base/optimization/constrained-kkt-qp-sqp-first-principles.md, 10-knowledge-base/optimization/objective-residual-design-and-audit.md, 10-knowledge-base/optimization/solver-selection-and-convergence-diagnosis.md, 10-knowledge-base/optimization/nonlinear-least-squares-first-principles.md, 10-knowledge-base/optimization/gauss-newton-levenberg-marquardt-dogleg.md, 10-knowledge-base/optimization/trust-region-line-search-globalization.md, 10-knowledge-base/optimization/jacobians-autodiff-manifold-linearization.md, 10-knowledge-base/optimization/factor-graph-solver-patterns-ceres-gtsam-g2o.md — residuals, whitening, constrained KKT/QP/SQP mechanics, Gauss-Newton, LM, dogleg, globalization, autodiff, manifold linearization, solver-library tradeoffs, and solver-failure triage |
| Numerical linear algebra | Sparse Estimation Backend Crosswalk, Nonlinear Solver Diagnostics Crosswalk, 10-knowledge-base/numerical-linear-algebra/cholesky-ldlt-normal-equations.md, 10-knowledge-base/numerical-linear-algebra/qr-svd-rank-revealing-solvers.md, 10-knowledge-base/numerical-linear-algebra/eigenvalues-hessian-conditioning-observability.md, 10-knowledge-base/numerical-linear-algebra/sparse-matrices-fill-in-ordering.md, 10-knowledge-base/numerical-linear-algebra/square-root-information-and-covariance-recovery.md, 10-knowledge-base/numerical-linear-algebra/schur-complement-marginalization-pcg.md — Cholesky/LDLT, QR/SVD, rank, nullspaces, sparse fill-in, orderings, square-root information, Schur complements, marginalization, PCG, and backend diagnostics |
| Geometry and mapping foundations | 10-knowledge-base/geometry-3d/lie-groups-se3-so3-jacobians.md, 10-knowledge-base/geometry-3d/camera-projective-geometry-pnp-triangulation.md, 10-knowledge-base/geometry-3d/fiducial-corner-localization.md, 10-knowledge-base/geometry-3d/epipolar-geometry-homographies-two-view.md, 10-knowledge-base/geometry-3d/optical-flow-scene-flow-first-principles.md, 10-knowledge-base/geometry-3d/point-cloud-registration-math-icp-ndt-gicp.md, 10-knowledge-base/geometry-3d/correspondence-search-data-structures.md, 10-knowledge-base/mapping/occupancy-bayes-evidential-dynamic-grids.md, 10-knowledge-base/geometry-3d/geodesy-map-projections-datums.md — Lie groups, projective geometry, PnP, fiducial/corner localization, triangulation, two-view epipolar/homography verification, optical/scene-flow motion fields, ICP/GICP/NDT, correspondence search, occupancy Bayes updates, and geodesy |
| Association, filters, and signals | 10-knowledge-base/state-estimation/data-association-and-gating.md, 10-knowledge-base/state-estimation/probabilistic-multi-object-association.md, 10-knowledge-base/state-estimation/information-filters-and-smoothers.md, 10-knowledge-base/state-estimation/particle-filters-and-hypothesis-management.md, 10-knowledge-base/sensors/sensor-likelihoods-noise-error-budgets.md, 10-knowledge-base/signal-processing/sampling-fft-windowing-filtering.md, 10-knowledge-base/signal-processing/radar-ambiguity-chirp-design-doppler-limits.md, 10-knowledge-base/signal-processing/cfar-detection-thresholding.md, 10-knowledge-base/signal-processing/sensor-filtering-alpha-beta-kalman-complementary.md, 10-knowledge-base/systems-engineering/time-sync-ptp-timestamping-latency-models.md, 10-knowledge-base/systems-engineering/benchmarking-metrics-statistical-validity.md — assignment, JPDA/MHT/RFS, information filters, particle filters, sensor likelihoods, FFT/filtering, radar ambiguity, CFAR, simple filters, timestamping, and statistical validity |
| Sensor measurement models | 10-knowledge-base/geometry-3d/lidar-working-principles-noise-models.md, 10-knowledge-base/geometry-3d/camera-imaging-noise-calibration.md, 10-knowledge-base/signal-processing/radar-fmcw-mimo-doppler.md, 10-knowledge-base/sensors/thermal-ir-radiometry-first-principles.md, 10-knowledge-base/sensors/ultrasonic-proximity-sensing-models.md — LiDAR, camera, radar, thermal IR radiometry, and ultrasonic acoustic ToF physics, noise, covariance, and calibration implications |
| IMU, GNSS, and wheel odometry | 10-knowledge-base/state-estimation/imu-error-models-preintegration.md, 10-knowledge-base/state-estimation/gnss-rtk-error-models.md, 10-knowledge-base/state-estimation/wheel-odometry-encoder-models.md — propagation, preintegration, RTK factors, dead reckoning, covariance, and outage behavior |
| Timing and calibration observability | 10-knowledge-base/systems-engineering/time-synchronization-error-budgets.md, 10-knowledge-base/geometry-3d/multi-sensor-calibration-observability.md, 10-knowledge-base/geometry-3d/active-calibration-experiment-design.md, 10-knowledge-base/geometry-3d/fiducial-corner-localization.md — timestamp error budgets, PTP/PPS, hand-eye calibration, observability motions, active/optimal calibration experiment design, fiducial/corner pose evidence, and online health checks |
| Event and thermal cameras | 10-knowledge-base/geometry-3d/event-thermal-camera-models.md, 10-knowledge-base/sensors/thermal-ir-radiometry-first-principles.md — event camera contrast model, timestamp noise, standalone thermal radiometry, NUC, and low-light/perception transfer |
| Lanelet2 | 10-knowledge-base/robotics/lanelet2-maps.md — airport extensions, AIXM conversion |
| Frenet planning | 10-knowledge-base/controls/frenet-trajectory-math.md — Werling 2010, quintic polynomials |
| Constrained control and belief-space decision-making | 10-knowledge-base/controls/constrained-optimization-mpc-ilqr-first-principles.md, 10-knowledge-base/controls/mdp-pomdp-belief-space-rl-first-principles.md — KKT conditions, MPC, iLQR, MDPs, POMDPs, belief states, and RL interfaces for learned autonomy |
| RTK/GPS/IMU | 10-knowledge-base/state-estimation/rtk-gps-imu-localization.md — preintegration, NTRIP |
| Mamba SSM | 10-knowledge-base/machine-learning/mamba-ssm-for-driving.md — DriveMamba, O(n) vs O(n²) |
| Theory | 10-knowledge-base/systems-engineering/theoretical-foundations.md — POMDP, free energy, PAC bounds |
| Architecture | 10-knowledge-base/systems-engineering/architecture-innovations.md — MoE, DiT, flow matching, FSQ |
| Sparse attention for 3D | 10-knowledge-base/machine-learning/sparse-attention-3d-perception.md — PTv3 serialized attention (80.4% mIoU, 3x faster), FlatFormer flattened windows (4.6x faster than SST), LitePT (CVPR 2026, 3.6x fewer params), SparseOcc, deformable attention, FlashAttention on Orin, TensorRT custom ops, hybrid SpConv+attention, multi-LiDAR cross-attention, window size 256-512 optimal for Orin |
Runtime, fleet, and validation topics
| Topic | Primary | Supporting |
|---|---|---|
| Sensor fusion | 30-autonomy-stack/perception/overview/sensor-fusion-architectures.md | |
| Synthetic data | 50-cloud-fleet/data-platform/synthetic-data-generation.md | |
| Airport-FOD3S synthetic FOD data engine | 50-cloud-fleet/data-platform/airport-fod3s-synthetic-data.md | Rare FOD synthesis workflow: FOD-A seed data, size/seam/style-controlled Airport-FOD3S composites, DualFOD/FOD-UAS and RDD5000 routing, LIDAROC proxy caveats, synthetic lineage, and real-only validation gates |
| Evaluation benchmarks | 60-safety-validation/verification-validation/evaluation-benchmarks.md | |
| nuScenes/Waymo guide | 30-autonomy-stack/perception/datasets-benchmarks/nuscenes-waymo-practical-guide.md | |
| Transfer learning | 50-cloud-fleet/mlops/transfer-learning.md | |
| ROS 2 migration | 40-runtime-systems/ros-autoware/ros2-migration.md | |
| Autoware Universe | 40-runtime-systems/ros-autoware/autoware-universe-deep-dive.md | |
| Open-source ecosystem | 40-runtime-systems/ml-deployment/opensource-ecosystem.md | |
| Embodied AI crossover | 10-knowledge-base/robotics/embodied-ai-crossover.md | |
| Data engine from bags | 50-cloud-fleet/data-platform/data-engine-from-bags.md | |
| Continual learning | 50-cloud-fleet/mlops/continual-learning.md | |
| 3D annotation tools | 50-cloud-fleet/data-platform/3d-annotation-tools.md | Open-source and commercial 3D annotation workflows, auto-labeling tiers, map-scale semantic candidate review, and MLOps scale controls for schema locks, QA, audit trails, vendor metrics, and safety-evidence promotion |
| Active labeling budget ops | 50-cloud-fleet/data-platform/active-labeling-budget-ops.md | Budgeted upload, auto-label inference, human annotation, expert review, QA, semantic-map candidate batches, taxonomy promotion, and S0-S5 label operations controls |
| Isaac ROS for airside | 40-runtime-systems/ros-autoware/isaac-ros-for-airside.md | |
| Test-time adaptation | 30-autonomy-stack/perception/overview/test-time-adaptation-airside.md | TENT, CoTTA, SAR, SFDA, OOD detection, active learning, multi-airport deployment |
| Test-time training for airport onboarding | 30-autonomy-stack/perception/overview/test-time-training-airport-onboarding.md | TTT vs TTA distinction (gradient-based auxiliary tasks), TTT++ multi-head, TTT-MAE, TTT layers as RNN, online LoRA with MAE loss, LiDAR-specific TTT (point cloud MAE), safety-bounded TTT on Orin (compute budget), catastrophic forgetting prevention (EWC, anchor loss), Simplex integration (TTT as AC, frozen as BC), airport onboarding protocol, comparison with PointLoRA, $25-45K/10-14 weeks |
| Fleet data pipeline | 50-cloud-fleet/data-platform/fleet-data-pipeline.md | RosBag/MCAP management, DVC/object snapshots, data product states, MLOps-scale orchestration, labeling workflows, fleet telemetry, retention, and storage costs |
| Data catalog, lineage, and quality operations | 50-cloud-fleet/data-platform/data-catalog-lineage-quality-ops.md | Data-product contracts, catalog/lakehouse/version-control architecture options, OpenLineage-style event boundaries, quality gate severity, snapshot identity, SLOs, deletion propagation, promotion states, release blockers, and safety-evidence retention |
| MLOps scale research scope | 50-cloud-fleet/mlops/mlops-scale-research-scope.md | MLOps maturity from notebook research to repeatable prototypes, production products, fleet/multi-site autonomy, regulated safety-critical release, and foundation-model/platform scale, covering data contracts, lineage, split/leakage firewalls, label operations, orchestration, registries, evaluation, serving, monitoring, governance, and autonomy-specific map/model/calibration evidence |
| MLOps reference architectures by scale | 50-cloud-fleet/mlops/mlops-reference-architectures-by-scale.md | Concrete S0-S5 MLOps architecture blueprints: minimum viable stack, centralize/decentralize/delay decisions, durable interfaces, fleet and regulated reference lanes, migration sequence, and platform-scale failure modes |
| MLOps migration checklist by scale | 50-cloud-fleet/mlops/mlops-migration-checklist-by-scale.md | Transition gates for S0->S1, S1->S2, S2->S3, S3->S4, and S4->S5, covering migration principles, entry/exit criteria, workstream migration matrix, tooling triggers, 30/60/90 plan, managed-site notes, and migration evidence packets |
| MLOps scorecards and KPIs by scale | 50-cloud-fleet/mlops/mlops-scorecards-and-kpis-by-scale.md | Scale-specific MLOps scorecards covering reproducibility, data and label quality, model/runtime quality, release reliability, observability, incident response, governance, cost, release-blocking metrics, cadence, ownership, and anti-metrics |
| Experiment tracking and reproducibility by scale | 50-cloud-fleet/mlops/experiment-tracking-reproducibility-by-scale.md | Run authority states, reproducibility levels, run manifest fields, tracker architecture options, comparison rules, and autonomy-specific LiDAR/image/map/ML-SLAM lineage controls across S0-S5 |
| Model registry and artifact lifecycle by scale | 50-cloud-fleet/mlops/model-registry-artifact-lifecycle-by-scale.md | Registry lifecycle guide for S0-S5 MLOps covering model/runtime/map/prompt/eval/replay artifact identity, alias authority, lifecycle states, artifact-set records, rollback retention, registry architecture options, semantic-map rules, and platform audit APIs |
| Serving and inference operations by scale | 50-cloud-fleet/mlops/serving-inference-operations-by-scale.md | Serving guide for S0-S5 MLOps covering batch inference, online endpoints, edge runtime packages, Triton/KServe/Seldon/Ray/BentoML/managed endpoint tradeoffs, inference service manifests, traffic routing, autoscaling, ODD-cell canaries, rollback, and serving observability |
| MLOps platform SRE and reliability by scale | 50-cloud-fleet/mlops/platform-sre-reliability-by-scale.md | Platform SRE guide for S0-S5 MLOps covering registry/eval/serving/orchestration reliability, criticality tiers, SLIs/SLOs, error budgets, backup/restore, DR, tenant isolation, incident lanes, audit logs, and platform bypass controls |
| Pipeline orchestration and release workflows by scale | 50-cloud-fleet/mlops/pipeline-orchestration-release-workflows-by-scale.md | Orchestrator selection and workflow state-machine guide for scripts, DVC, GitHub Actions, Airflow, Argo, Kubeflow, TFX, Ray, Slurm, and managed ML pipelines, covering artifact handoffs, release/evidence gates, incident lanes, managed-site rules, and S0-S5 scorecards |
| Evaluation platforms and replay gates by scale | 50-cloud-fleet/mlops/evaluation-platform-replay-gates-by-scale.md | Evaluation-platform guide for S0-S5 MLOps covering metric specs, evaluator identity, evaluation manifests, replay packages, runtime package checks, shadow/canary evidence, safety waivers, ODD-cell release blockers, LiDAR/image semantic-map evaluation, training-architecture coupling, and platform SLOs |
| Dataset split and leakage controls by scale | 50-cloud-fleet/mlops/dataset-split-leakage-controls-by-scale.md | Split-firewall guide for S0-S5 MLOps covering split manifests, temporal/site/route/vehicle/map/labeler/feature/synthetic/federated leakage, LiDAR-image and ML-SLAM-specific controls, non-road managed-site holdouts, training architecture tradeoffs, scorecards, and release blockers |
| Model monitoring and drift response by scale | 50-cloud-fleet/mlops/model-monitoring-drift-response-by-scale.md | Monitoring and drift-response guide for S0-S5 MLOps covering service health, data quality, training-serving skew, prediction drift, delayed labels, map/calibration drift, alert routing, retraining trigger policy, ODD-cell quarantine, rollback, safety-case deltas, and platform alert-quality controls |
| Site-sliced release evidence by scale | 50-cloud-fleet/mlops/site-sliced-release-evidence-by-scale.md | ODD-cell release evidence guide for site, route, task, weather, vehicle kit, map/calibration state, artifact set, local holdouts, replay, shadow, canary, delayed-label review, safety-case linkage, and rollback |
| Feature and embedding store operations by scale | 50-cloud-fleet/mlops/feature-embedding-store-ops-by-scale.md | Store-selection and control guide for manifest-backed files, offline/online feature stores, lakehouse tables, and embedding/vector stores across S0-S5, covering point-in-time leakage, offline/online skew, vector-index reproducibility, retrieval recall, deletion propagation, and autonomy-specific use cases |
| Offboard labeler registry by scale | 50-cloud-fleet/mlops/offboard-labeler-registry-by-scale.md | Registry pattern for heavy offline labelers, prompt packs, open-vocabulary 3D labelers, map-derived exporters, weak rules, LLM/VLM reviewers, evaluator models, retrieval corpora, thresholds, allowed-use states, and reviewer workflows across S0-S5 |
| GPU queueing and FinOps by scale | 50-cloud-fleet/mlops/gpu-queueing-finops-by-scale.md | GPU workload classes, queue architecture, scheduler choices, priority lanes, per-job metadata, capacity planning, unit economics, safety-evidence capacity, monitoring, and S0-S5 FinOps controls for training, replay, labeling, map segmentation, and incidents |
| Secure artifact attestation by scale | 50-cloud-fleet/mlops/secure-artifact-attestation-profile.md | Trust-chain profile for signed containers, model weights, ONNX/TensorRT engines, semantic maps, map-hygiene layers, dataset manifests, prompt packs, label batches, replay/eval packs, SBOMs, SLSA/in-toto provenance, registry aliases, and policy verification across S0-S5 |
| Federated and privacy-preserving training policy | 50-cloud-fleet/mlops/federated-privacy-preserving-training-policy-by-scale.md | Trigger policy for centralized, local-only, hybrid, federated, secure-aggregation, differential-privacy, and confidential-compute training decisions across S0-S5, with architecture comparisons, client contracts, privacy controls, evaluation gates, and managed-site rules |
| LLMOps and agent evaluation by scale | 50-cloud-fleet/mlops/llmops-agent-evaluation-by-scale.md | GenAIOps/LLMOps controls for prompt packs, model endpoints, RAG corpora, tool policies, agent graphs, judge models, eval packs, traces, guardrails, release gates, telemetry, and managed-site safety boundaries across S0-S5 |
| Map-derived pseudo-label invalidation | 50-cloud-fleet/mlops/map-derived-pseudo-label-invalidation-protocol.md | Invalidation protocol for semantic-map-derived labels: triggers, state machine, impact graph, batch manifest fields, S0-S5 controls, and airside release-state rules before labels feed training, replay, eval, or safety evidence |
| Data flywheel (closed-loop) | 50-cloud-fleet/mlops/data-flywheel-airside.md | Trigger-based collection (50GB/day budget), auto-labeling (70-85% cost reduction), map-derived semantic labels gated by release-state eligibility, active learning (40-50% fewer labels), model training orchestration, A/B testing, scenario mining, multi-airport LoRA, flywheel breakeven ~Month 18, mAP trajectory 45%→82% over 24 months |
| Radar-LiDAR fusion for adverse weather | 30-autonomy-stack/perception/overview/radar-lidar-fusion-adverse-weather.md | L4DR (AAAI 2025, +20% mAP dense fog), Continental ARS548 4D radar ($500-1500), asymmetric mid-level fusion (LiDAR-primary, radar-augmented), radar-guided densification, cross-attention LiDAR→radar, adaptive fusion gating (weather-aware weights), de-icing spray detection, track-level Kalman fusion, 4-mode degradation management (NORMAL→EMERGENCY), ROS integration, $35-55K/12 weeks |
| Federated learning (fleet) | 50-cloud-fleet/mlops/federated-learning-fleet.md | FedAvg/FedProx/SCAFFOLD, hybrid centralized+federated LoRA (97% comm reduction), FedBN for multi-airport, on-vehicle Orin LoRA training, DP privacy (epsilon 10-50), Byzantine-robust FLTrust, federated continual learning, hierarchical aggregation, FedDF for heterogeneous models, Flower/FLARE, break-even ~10 airports, $130K/year at 50 airports vs $1.3M centralized |
| LiDAR data augmentation | 50-cloud-fleet/mlops/lidar-data-augmentation.md | GT-database sampling (+15-25% rare class AP), 3D copy-paste, PolarMix (CVPR 2022, +3-7% mAP), LaserMix (CVPR 2023), LiDAR corruptions (rain/fog/beam dropout/de-icing), intensity augmentation, class-balanced sampling with safety priority, cross-airport GT database sharing, 40-60% labeling reduction ($15-45K savings/airport) |
| Cloud backend infrastructure | 50-cloud-fleet/data-platform/cloud-backend-infrastructure.md | Fleet data backend: three-zone data lake (Raw/Bronze → Processed/Silver → Curated/Gold), S3 event-driven ingestion (Lambda), streaming telemetry (MQTT→IoT Core→TimeStream→Grafana), Apache Airflow DAG catalog (7 DAGs), rosbag processing K8s jobs, feature store (Feast), auto-labeling pipeline integration, map construction data flow, multi-airport data isolation, cost modeling ($200-460/vehicle/month), monitoring/observability, $80-135K/28 weeks |
| On-vehicle data triage & upload | 40-runtime-systems/data-logging/on-vehicle-data-triage-selective-upload.md | Vehicle-side data management: multi-tier ring buffers (LiDAR/camera/IMU/CAN/GTSAM, NVMe 1-4TB), event-triggered clip extraction (safety events, perception anomalies, localization failures, operator flags), edge scenario classification (lightweight CNN/DLA), bandwidth-aware upload scheduling (priority queue, 50GB/day budget), compression (LZ4 point clouds, H.265 camera, delta poses), rosbag split/trim/mcap, active learning integration, fleet upload coordination (deduplication, coverage diversity), GDPR camera data handling, ROS node architecture |
| Sensor calibration fleet operations | 40-runtime-systems/software-operations/sensor-calibration-fleet-ops.md | Calibration package lifecycle, manifest fields, versioned telemetry, drift remediation, quarantine/rollback, release gates, maintenance recovery, and calibration evidence artifacts |
Synthesis & strategy
| Topic | Primary |
|---|---|
| Master synthesis | 90-synthesis/master/master-synthesis.md — Executive summary, tiered recommendations |
| Design spec | 90-synthesis/decisions/design-spec.md — 891-line Simplex architecture |
| POC proposals | 90-synthesis/poc-roadmaps/poc-proposals.md — 8 models with code and costs |
| Competitive landscape | 80-industry-intel/market-competitive/competitive-landscape.md — All players compared, strategic quadrant |
| Technology readiness | 90-synthesis/readiness-risk/technology-readiness.md — TRL per POC, go/no-go criteria |
| Knowledge gap backlog | 90-synthesis/readiness-risk/knowledge-gap-backlog.md — P0/P1/P2 missing research files across the end-to-end AV architecture |
| Active frontier source registry | 90-synthesis/readiness-risk/active-frontier-source-registry.md — Manual source and query registry for perception, SLAM, world models, VLA/VLM, datasets, and validation monitoring |
| Getting started | 90-synthesis/master/getting-started.md — Day 1 guide with runnable code |
Recently Added (Latest Sessions)
| Document | Key Contribution |
|---|---|
| Recurring RKO-LIO batch | RKO-LIO was promoted as a sensor-agnostic LiDAR-inertial odometry page covering simplified IMU use, scan-to-map regularization, ROS 2/Python packaging, deployment assumptions, and cross-domain failure modes |
20-av-platform/sensors/sensor-to-algorithm-readiness-contract.md | Bridge contract for sensor acquisition, calibration, synchronization, preprocessing, health, provenance, and algorithm input acceptance before perception/SLAM/fusion consumers run |
20-av-platform/sensors/calibration-bay-fixtures.md | Physical calibration-bay fixtures and evidence: bay reference frame, target/fixture inventory, lighting/thermal/radar controls, capture manifest, signed package handoff, and airside post-maintenance checks |
90-synthesis/readiness-risk/active-frontier-source-registry.md | Manual-first registry of active frontier sources, native filters, query patterns, canonical routing rules, and semi-automation boundaries |
| Web gap expansion wave | 31 source-backed files covering 4D radar-camera occupancy, CVFusion, FMCW LiDAR predictive detection, cross-domain scene flow, TrackOcc, dynamic 3DGS/4DGS, DistillNeRF, self-supervised occupancy flow, UniScene, robust/certifiable SLAM backends, lifelong map maintenance, GPR/radar localization, probability/control foundations, adverse/OOD/FOD datasets, and validation protocols |
| Perception/SLAM reliability gap wave | 36 source-backed files covering occupancy fusion, dynamic/free-space occupancy, radar-LiDAR adverse-weather detection, RobuRCDet, SAMFusion, STU, synthetic FOD, OVAD/OVODA, open-vocabulary panoptic occupancy, RCP-Bench, V2X sequential datasets, Scan Context, LiDAR BA factors, Kimera-Multi, COVINS/COVINS-G, D2SLAM, UWB/range SLAM, OKVIS2-X, MM-LINS, event/thermal/radar localization, continuous-time and volumetric-map foundations, detection/tracking foundations, fleet-data contracts, and perception/SLAM/map validation protocols |
| First-principles foundations wave | 37 source-backed KB files covering Gaussian noise, Mahalanobis/chi-square gating, MAP/MLE, robust statistics, mixtures, Gauss-Newton, LM, dogleg, Jacobians, residual audits, solver convergence diagnosis, Ceres/GTSAM/g2o, Cholesky/LDLT, QR/SVD, sparse solvers and backend crosswalks, square-root information, Schur/PCG, Lie groups, projective geometry, ICP/GICP/NDT, occupancy grids, geodesy, assignment, JPDA/MHT/RFS, filters, sensor likelihoods, signal processing, radar ambiguity, CFAR, timestamping, and statistical benchmarking |
90-synthesis/readiness-risk/continuous-research-loop.md | Continuous research loop for discovery, triage, atomic-file promotion, cross-linking, verification, and next-queue selection across perception, SLAM, sensors, and mapping |
| Recurring Gaussian/foundation SLAM batch | HI-SLAM2, SEGS-SLAM, and Neural/Gaussian SLAM Surveys were promoted as current visual 3DGS SLAM method and taxonomy pages, with T-RO, ICCV, project, repository, and survey sources checked |
| Recurring radar/SLAM hardening batch | GV-iRIOM was promoted as a globally referenced 4D radar/visual/GNSS mapping page, with radar-inertial online calibration refreshed for LC-RIO-ET-style spatio-temporal calibration |
| Recurring learned 4D radar odometry batch | CAO-RONet was promoted as a learned 4D radar odometry page covering local completion, context-aware association, clip-window optimization, repository maturity, View-of-Delft evaluation context, and adverse-weather transfer caveats |
| Recurring Doppler radar bridge batch | Doppler Radar-LiDAR SLAM was promoted as a bridge page for Radarize, DRO, and Doppler-SLAM, covering Doppler-shift odometry, direct radar registration, Doppler-aided radar/LiDAR/inertial fusion, code maturity caveats, and cross-domain failure modes |
| Recurring radar RIO correspondence/uncertainty batch | Radar RIO Correspondence and Uncertainty was promoted as a combined radar-inertial hardening page for learned point correspondences, polar point uncertainty, continuous point-pose uncertainty, UNRIO routing, and AV fallback validation caveats |
| Recurring radar place-recognition refresh | Radar Place Recognition was refreshed as a 4D radar descriptor lineage page covering 4DRaL, SHeRLoc, TransLoc4D, 4D RadarPR, TDFANet, DIDLM dataset routing, code-maturity caveats, and verification boundaries |
| Recurring Gaussian occupancy batch | GaussianFlowOcc and GaussTR were promoted as source-mature sparse/foundation-aligned Gaussian occupancy method pages, and LinkOcc source routing was corrected away from the RobuRCDet arXiv ID |
| Recurring GS-Occ3D occupancy batch | GS-Occ3D was promoted as a vision-only Gaussian-surfel occupancy reconstruction and binary-label curation page, with LinkOcc kept on watchlist pending open method details or code |
| Recurring SLAM benchmark matrix refresh | Oxford Spires, IILABS 3D, SMapper-light, FusionPortableV2, and S3E were routed into the SLAM benchmark matrix with source caveats and no new atomic files |
| Recurring audit/navigation cleanup | Count drift was corrected for companies, perception methods, and airside operations; duplicate-prone active queue wording was tightened around EvOcc, DR-REMOVER, ExelMap, and already-promoted perception pages |
| Recurring perception/sparse/occupancy/dataset batch | DepthOcc, SparseBEV, DETR4D, TEOcc, GaussRender, DSERT-RoLL, and CMHT were promoted across camera occupancy, sparse-query detection, temporal occupancy, Gaussian supervision, and multimodal adverse-weather dataset coverage |
| Recurring research loop promoted batch | SLAM Toolbox and NDT variants/NDT maps were promoted as first-class SLAM method pages, with ROS 2 package, Autoware NDT, PCL NDT, fast_gicp, and ndt_omp source links checked |
| Recurring visual/VIO benchmark batch | LaMAria, Hilti x Trimble SLAM Challenge 2026, and ScaleMaster were promoted as current SLAM benchmark pages for egocentric VIO, construction floor-plan localization, and monocular scale/map-quality testing |
| Recurring truck-centered V2X dataset batch | TruckV2X was promoted as a truck-centered cooperative perception dataset page covering tractor/trailer/CAV/RSU agents, synthetic scenarios, occlusion recovery analysis, and articulated-vehicle transfer caveats |
| Recurring collaborative Gaussian occupancy batch | VOGS-CP was promoted as a collaborative Gaussian semantic occupancy method covering sparse 3D semantic Gaussian messages, ROI-culling, cross-agent Gaussian fusion, Gaussian-to-voxel splatting, and simulation-to-real transfer caveats |
| Recurring FOD synthetic data-engine batch | Airport-FOD3S synthetic FOD data engine was promoted for rare FOD generation, physical-size transformation, seam/style controls, synthetic lineage, real-only validation gates, DualFOD/FOD-UAS routing, and LIDAROC sensor-contamination proxy caveats |
| Recurring raw GNSS factor fusion batch | GVINS/GLIO was promoted as a raw GNSS factor-fusion page covering pseudorange and Doppler factors, visual-inertial and LiDAR-inertial graph companions, urban-canyon caveats, and validation gates |
| Recurring wheel/vehicle-motion factor batch | Wheel odometry and vehicle-motion factors were promoted as a SLAM/localization factor-family page covering visual-inertial-wheel odometry, nonholonomic constraints, skid-steer calibration, online neural wheel-kinematic factors, slip caveats, and validation gates |
| Recurring embodied 3D benchmark batch | EmbodiedScan/MMScan was promoted as an embodied 3D benchmark page covering egocentric RGB-D scene understanding, 3D boxes, semantic occupancy, hierarchical grounded language annotations, visual grounding, 3D QA, and AV/VLM transfer caveats |
| Recurring thermal radiometry batch | Thermal IR Radiometry First Principles was promoted as a sensor-foundation page covering emitted radiance, reflected background, atmospheric path, microbolometer response, NUC tables, radiometric parameters, and fusion-ready confidence |
| Recurring fleet calibration operations refresh | Sensor Calibration Fleet Operations was refreshed with package lifecycle states, artifact manifest fields, versioned telemetry schemas, drift remediation workflow, anti-patterns, rollback/quarantine gates, and current source links |
| Recurring infrastructure-aided localization batch | Infrastructure-Aided Localization was promoted as a localization overview covering UWB, fiducials, RFID/BLE, Wi-Fi RTT, magnetic maps, reflectors, 5G NR/mmWave positioning, measurement contracts, lifecycle controls, and estimator handoff patterns |
| Recurring fiducial/corner localization batch | Fiducial and Corner Localization was promoted as a geometry foundation page covering AprilTag, ArUco, ChArUco, checkerboard and AprilGrid targets, planar PnP/IPPE pose evidence, marker-map survey contracts, estimator handoff, and managed-site failure modes |
| Recurring semantic-map navigation cleanup | Post-merge corpus counts, perception audit counts, and README/INDEX routes were synchronized for the aggregated-map semantic segmentation hub and companion pages covering taxonomy, training, tiling, post-processing, and static-transient removal |
| Recurring ML-SLAM semantic-map scope loop | ML-related SLAM was promoted as a cross-section research-scope page for learned registration, learned place recognition, semantic/dynamic SLAM, neural/Gaussian maps, dynamic residual removal, static-but-transient quarantine, map priors, and aggregated-map segmentation handoffs |
| Recurring semantic-map first-principles loop | Semantic mapping and point-cloud segmentation KB pages were deepened with permanence as a separate state variable, semantic-layer production, map-hygiene metrics, false-permanent/false-deletion rates, and transient-label leakage controls |
| Recurring semantic-map dataset/training loop | Large-scale 3D segmentation benchmarks gained a release-oriented dataset-selection protocol, and training paradigms gained a map-scale backbone-by-supervision decision matrix for sparse-conv, KPConv/RandLA, SPT, PTv3/Sonata, projection, LiDAR-image, and SSM/Mamba routes |
| Recurring semantic-map removal loop | Dynamic-removal, static-transient, benchmark, and ML-SLAM scope pages gained layered removal gates for permanent static, dynamic residual, static transient, movable-static, FOD candidate, artifact, and review-required map points |
| Recurring semantic-map release-contract loop | Semantic-map manifest schema and examples now hash-address map-hygiene layers and require hygiene metric vectors, and map-publication gates block release when those machine-checkable outputs are missing |
| Recurring semantic-map modality-contract loop | Camera-LiDAR fusion was expanded into a map-modality contract covering LiDAR-only, pre-baked colorized clouds, train-time image distillation, image-dependent fusion, candidate-label lanes, and projection QA evidence |
| Recurring semantic-map taxonomy/permanence loop | Taxonomy design now separates semantic class IDs from map-hygiene permanence layers, with non-road urban-district mappings for apron, campus, port, industrial, utility, and facade domains plus manifest-backed release rules |
| Recurring semantic-map tiling/postprocess loop | Tiling and post-processing companions now define tile release ledgers plus semantic, confidence, and map-hygiene output layers that connect per-tile logits to manifest-backed publication evidence |
| Recurring semantic-map source-handoff loop | Map construction, MapEval, publication gates, and the aggregated-map hub now define a source-map acceptance package that gates semantic segmentation on geometry, calibration, hygiene layers, projection QA, and quarantine decisions |
| Recurring semantic-map removal-decision loop | Static-transient and detector-ground-projection pages now require reason-coded release-state outputs and rejected-object evidence sidecars for stationary people, parked assets, FOD candidates, artifacts, and unknown-review points |
| Recurring semantic-map ML-SLAM handoff loop | ML-related SLAM and learned-prior map pages now separate learned evidence that may propose, smooth, rank, or route review from observed source-map evidence that can publish release truth |
| Recurring semantic-map release-state benchmark loop | MOS and dynamic-map-cleaning benchmark pages now define release-state labels and confusion metrics for permanent-static, dynamic-residual, movable-static, static-transient, FOD-candidate, artifact, and unknown-review points |
| Recurring semantic-map training-export loop | Training, taxonomy, ground-truth, and publication-gate pages now require map-derived auto-labels to carry semantic class plus release-state labels, with permanent_static as the only default supervised positive and all transient/hazard/artifact/unknown states masked, quarantined, or routed to active learning |
| Recurring semantic-map seam/postprocess loop | Tiling, post-processing, monitoring, and hub pages now require release-state seam confusion, release-state-preserving smoothing, and runtime alerts for false-permanent drift or contaminated training exports |
| Recurring semantic-map flywheel-governance loop | Data flywheel, single-scan segmentation, and model-governance pages now carry semantic-map release-state eligibility into pseudo-label promotion, training manifests, and model release evidence |
| Recurring semantic-map schema-contract loop | Semantic-map manifest schema, example, tests, and hub routing now make training-export release-state eligibility machine-checkable before pseudo-label batches are consumed |
| Recurring semantic-map pseudo-label invalidation loop | MLOps scale, data catalog, data flywheel, model governance, semantic-map hub, README, and INDEX now define a state machine and impact graph for invalidating map-derived pseudo-label batches after source-map, calibration, taxonomy, cleaner, reviewer, prompt, split, privacy, or incident triggers |
| Recurring site-sliced release evidence loop | MLOps scale, model governance, scorecards, reference architecture, production deployment, shadow mode, ODD monitoring, README, and INDEX now define ODD-cell release manifests, local holdouts, shadow/canary evidence, delayed-label review, map-state scope, waiver expiry, and rollback gates so one aggregate score cannot approve every managed site |
| Recurring feature/embedding store MLOps loop | MLOps scale, reference architecture, scorecards, data catalog, scenario mining, model governance, semantic-map pipeline, README, and INDEX now define when to use manifests, offline feature tables, online feature stores, lakehouse tables, and vector stores, including leakage, skew, index recall, deletion propagation, and safety-evidence controls |
| Recurring offboard labeler registry loop | MLOps scale, model governance, data flywheel, active labeling, annotation tools, semantic-map hub, README, and INDEX now treat prompt packs, foundation-model labelers, evaluator/judge models, weak label rules, retrieval corpora, thresholds, reviewer workflows, allowed-use states, and rollback bundles as governed release-affecting artifacts |
| Recurring GPU queueing/FinOps loop | MLOps scale, reference architecture, scorecards, training infrastructure, cloud backend, data flywheel, README, and INDEX now define GPU workload classes, queue priority lanes, scheduler choices, required job metadata, unit economics, reserved assurance capacity, and safety-aware FinOps controls across S0-S5 |
| Recurring secure artifact attestation loop | MLOps scale, reference architecture, scorecards, model governance, data catalog, training infrastructure, OTA/SUMS, compatibility matrix, README, and INDEX now define digest-bound signing, SBOM/provenance, trusted-builder evidence, registry alias policy, and deployment verification for models, maps, prompts, labels, eval packs, and containers |
| Recurring MLOps migration checklist loop | MLOps scale, reference architecture, scorecards, model governance, fleet data pipeline, training infrastructure, README, and INDEX now define S0-S5 migration gates, entry/exit criteria, workstream upgrade matrix, tooling triggers, 30/60/90 adoption plan, and migration evidence packets |
| Recurring dataset split/leakage controls loop | MLOps scale, scorecards, model governance, data catalog, fleet data pipeline, site-sliced evidence, pseudo-label invalidation, README, and INDEX now define split manifests and leakage controls for temporal logs, routes, sites, vehicles, map tiles, labelers, feature/vector stores, synthetic data, federated clients, LiDAR-image fusion, and ML-SLAM-derived labels |
| Recurring model monitoring/drift response loop | MLOps scale, reference architecture, scorecards, model governance, production deployment, fleet anomaly attribution, fleet SRE, README, and INDEX now define monitoring event contracts, drift signal taxonomy, retraining trigger policy, ODD-cell quarantine, rollback, safety-case deltas, suppression expiry, and alert-quality KPIs |
| Recurring experiment tracking/reproducibility loop | MLOps scale, reference architecture, migration checklist, scorecards, model governance, training infrastructure, README, and INDEX now define run authority states, reproducibility levels, tracker architecture choices, run manifest fields, comparable-baseline rules, and autonomy-specific LiDAR/image/map/ML-SLAM lineage controls |
| Recurring model-registry/artifact-lifecycle loop | MLOps scale, reference architecture, migration checklist, scorecards, model governance, attestation, compatibility, runtime deployment, README, and INDEX now define registry records, alias authority, lifecycle states, artifact-set membership, rollback retention, semantic-map artifact scope, and platform registry audit controls |
| Recurring serving/inference-operations loop | MLOps scale, reference architecture, migration checklist, scorecards, registry, evaluation, monitoring, runtime deployment, README, and INDEX now define batch/online/edge serving modes, inference service manifests, serving platform choices, traffic routing, autoscaling, ODD-cell rollout, rollback, and observability controls |
| Recurring MLOps platform-SRE/reliability loop | MLOps scale, reference architecture, migration checklist, scorecards, orchestration, registry, serving, evaluation, monitoring, FinOps, fleet SRE, README, and INDEX now define platform service criticality tiers, SLOs, error budgets, backup/restore, DR, tenant isolation, incident lanes, and platform bypass controls |
| Recurring pipeline orchestration/release-workflow loop | MLOps scale, reference architecture, migration checklist, scorecards, training infrastructure, fleet data pipeline, README, and INDEX now define workflow authority states, orchestrator choices, typed artifact handoffs, release state machines, incident/evidence lanes, queue coupling, and managed-site workflow gates |
| Recurring evaluation-platform/replay-gate loop | MLOps scale, reference architecture, migration checklist, scorecards, model governance, site-sliced evidence, scenario mining, README, and INDEX now define evaluation manifests, metric specs, evaluator identity, replay/runtime package gates, shadow/canary evidence, waiver controls, LiDAR/image semantic-map evaluation, and platform evaluation SLOs |
| Recurring federated/privacy training policy loop | MLOps scale, reference architecture, migration checklist, scorecards, model governance, data privacy, fleet data pipeline, federated learning guide, README, and INDEX now define when centralized, local, hybrid, federated, secure-aggregation, differential-privacy, or confidential-compute training is justified and how release evidence is gated |
| Recurring LLMOps/agent evaluation loop | MLOps scale, reference architecture, scorecards, model governance, offboard labeler registry, feature/embedding store, VLM scene understanding, README, and INDEX now define prompt, RAG, judge, VLM/VLA, tool-agent, trace, telemetry, prompt-injection, and release-gate controls across S0-S5 |
| Recurring semantic-map non-road benchmark-bundle loop | Large-scale benchmark, GridNet-HD, and aggregated-map hub pages now define non-road urban-district benchmark bundles and release-state overlay requirements for apron/depot, campus, port/industrial, utility, facade, and terminal-interior slices |
| Recurring semantic-map training-objective loop | Loss/metrics first principles, training paradigms, and the aggregated-map hub now define a release-state-aware multi-head objective so map-derived labels supervise semantic classes only when the source point is exportable permanent-static evidence |
| Recurring semantic-map removal-handoff loop | Dynamic-removal, static-transient, hygiene-protocol, hub, and README routes now require reason-coded removal sidecars with release-state aliases, evidence fields, review state, and downstream permissions for deleted or quarantined map points |
| Recurring semantic-map application-architecture loop | Aggregated-map hub, map-publication gates, data-flywheel, README, and INDEX now separate runtime semantic maps, training exports, hygiene monitoring, digital-twin transfer, and local benchmark acceptance as distinct product modes with separate evidence gates |
| Recurring semantic-map product-mode contract loop | Semantic-map manifest schema, example, tests, hub, and map-publication gates now make product-mode gates machine-checkable for runtime maps, training exports, hygiene monitoring, digital-twin transfer, and benchmark acceptance |
| Recurring semantic-map architecture-routing loop | Perception method library now routes aggregated-map segmentation architecture choices through sparse-conv production anchors, point-conv geometry baselines, superpoint graph map-scale context, serialized transformer/foundation-model ceilings, projection deployment lanes, image-distilled/open-vocabulary label lanes, and SSM/Mamba experiments with shared bake-off controls |
| Recurring semantic-map method-control loop | PTv3, WaffleIron, and 2DPASS method pages now carry release-map controls for accuracy-ceiling transformers, projection-based deployment lanes, and train-time image distillation, including product-mode permissions, projection/serialization manifests, release-state masks, and non-road transfer audits |
| Recurring semantic-map removal-stack loop | Dynamic-removal, static-transient, and LiDAR artifact-removal pages now separate artifact filtering, in-session dynamic residual removal, stationary-person exclusion, parked movable-asset quarantine, FOD-candidate handling, infrastructure-change review, and protected thin-structure preservation before semantic-map publication |
| Recurring semantic-map product-taxonomy loop | Taxonomy and benchmark pages now map runtime maps, training exports, hygiene monitoring, digital twins, and local acceptance sets to minimum class coverage, proxy bundles, owned-evidence requirements, and release-state gates |
| Recurring semantic-map ML-SLAM product-routing loop | ML-SLAM, map-construction, and learned-prior pages now route learned registration, place recognition, dynamic removal, semantic priors, neural/Gaussian maps, and back-projected labels by product mode so runtime maps, training exports, hygiene monitoring, digital twins, and benchmarks do not share unsafe evidence shortcuts |
| Recurring semantic-map maintenance-routing loop | Map maintenance, moved-object benchmarks, and publication gates now distinguish detected semantic changes from release actions so runtime patches, training-label expiry, hygiene tickets, digital-twin refreshes, and benchmark cases follow separate evidence paths |
| Recurring MLOps scale-scope loop | MLOps now has a scale ladder from notebook research through production, fleet autonomy, regulated safety-critical release, and foundation-model/platform scale, with companion updates to data flywheel, model governance, runtime deployment, README, and INDEX routing |
| Recurring MLOps implementation-scale loop | Fleet data, cloud backend, and training infrastructure pages now translate the MLOps scale ladder into concrete data-platform, backend-service, and GPU-training upgrade triggers from prototype through regulated multi-site fleet operation |
| Recurring MLOps monitoring-assurance loop | Runtime monitoring, ML assurance, and perception/SLAM runbooks now route drift, runtime, canary, incident, and safety-case signals by MLOps scale so prototype anomalies, production blockers, fleet ODD regressions, and regulated safety events get different evidence and escalation |
| Recurring MLOps foundation-model governance loop | MLOps, model-governance, VLM scene-understanding, and spatial-foundation-model pages now govern prompt packs, foundation-model checkpoints, retrieval corpora, evaluator/judge models, tool policies, traces, and VLM/VLA advisory boundaries by scale so generated labels and reasoning cannot silently become release truth |
| Recurring MLOps security-finops loop | MLOps, cybersecurity, and training-infrastructure pages now scale secure artifact attestation, SBOM/signing, site-scoped data access, privacy retention, GPU queues, quotas, utilization, and unit-cost evidence from notebooks through regulated platform operations |
| Recurring MLOps operating-model loop | MLOps scale, model-governance, and data-flywheel pages now define ownership, toolchain build/buy choices, release RACI, and cadence by scale so research, production, fleet, regulated, and platform teams know who owns each artifact and promotion decision |
| Recurring MLOps data-product loop | MLOps scale, data catalog, fleet data pipeline, reference architecture, migration, and scorecard pages now define data-product contracts, catalog/lakehouse/version-control choices, OpenLineage-style event boundaries, quality gate severity, feature/embedding store triggers, SLOs, deletion propagation, and safety/legal-hold boundaries from S0 notebooks through S5 platform operations |
| Recurring MLOps reference-architecture loop | MLOps scale, model-governance, data-flywheel, CI/CD, README, and INDEX now define concrete S0-S5 reference architectures, centralization boundaries, durable interfaces, migration sequence, and platform bypass failure modes |
| Recurring MLOps scorecard loop | MLOps scale, reference architectures, data flywheel, fleet observability, README, and INDEX now define S0-S5 scorecards, release-blocking metrics, operating cadence, ownership, managed-site KPI focus, and anti-metrics |
| Recurring MLOps compatibility loop | Artifact compatibility matrix, model governance, MLOps reference architecture, MLOps scorecards, semantic-map pipeline, and INDEX now scale compatibility evidence across model, map, calibration, runtime, telemetry, semantic taxonomy, prompt/labeler, replay, and rollback artifacts |
| Recurring GridNet-HD dataset batch | GridNet-HD was promoted as a LiDAR-image utility-infrastructure benchmark for thin pylon, cable, insulator, vegetation, soil/road, water, and building classes, with routing through the benchmark, taxonomy, hub, audit, source registry, README, INDEX, and methodology surfaces |
| Recurring ZAHA facade-hierarchy batch | ZAHA was routed as a WACV 2025 / TUM2TWIN MLS facade benchmark with 601M annotated points, LoFG2/LoFG3 hierarchy, and terminal-frontage/digital-twin taxonomy transfer notes |
| Recurring OpenLiDARMap georeferencing batch | OpenLiDARMap and FlexCloud were routed as source-backed georeferenced point-cloud map-conditioning branches for GCP-sparse or GNSS-assisted map construction, with explicit provenance/residual gates before semantic segmentation |
| Recurring LAMM multi-session map-merging batch | LAMM was promoted as a source-backed large-scale multi-session point-cloud map-merging page and routed as an upstream map-substrate conditioning branch before aggregated-map semantic segmentation |
| Recurring modality-aware architecture refresh | Aggregated-map segmentation now separates colorized clouds, train-time distillation, image-dependent fusion, and candidate-label lanes, then compares sparse-conv, KPConv/RandLA, SPT, PTv3/Sonata, projection, and SSM/Mamba backbones by LiDAR-only fit, image dependency, map partitioning, and production role |
| Recurring Uni-Mapper map-merging loop | Uni-Mapper was promoted as a dynamic-aware heterogeneous-LiDAR map-merging page and routed through the aggregated-map segmentation, static-transient, dynamic-cleaning, map construction, SLAM audit, stack comparison, README, and INDEX surfaces |
| Recurring LOSC consolidation batch | LOSC was promoted as an atomic perception method page for open-vocabulary LiDAR pseudo-label consolidation, then routed through aggregated-map segmentation, open-vocabulary detection, data-engine, annotation, audit, README, and INDEX surfaces |
| Recurring map-hygiene workflow cleanup | Canonical map-hygiene ground truth and its V&V companion were routed through aggregated-map QA, map publication gates, operational monitoring, regulatory evidence, README, INDEX, and the active loop queue |
| Perception/SLAM/sensor deep-dive wave | 33 source-backed files covering SplatAD and Gaussian/4DGS perception, latest sparse/radar-camera perception, production LIVO/SLAM, Gaussian/radar SLAM, and sensor measurement/noise fundamentals |
10-knowledge-base/, 20-av-platform/, 30-autonomy-stack/, 40-runtime-systems/, 50-cloud-fleet/, 60-safety-validation/, 70-operations-domains/ P0 gap wave | 35 source-backed P0 gap files covering foundations, platform power/diagnostics/ruggedization, planning/control/V2X, E2E/VLA/world models, runtime/cloud operations, safety evidence, and non-airside operations domains |
90-synthesis/readiness-risk/knowledge-gap-backlog.md | Cross-architecture gap backlog from parallel research agents: P0/P1/P2 missing files across foundations, platform, autonomy, runtime/cloud, safety, operations, and industry intelligence |
30-autonomy-stack/localization-mapping/overview/production-lidar-map-localization.md | Production scan-to-map matching: VGICP/NDT/ICP comparison, multi-resolution coarse-to-fine, eigenvalue degeneracy detection, multi-LiDAR fusion strategies, GTSAM adaptive noise, 5-level fallback, GeoTransformer cold start, 15-25ms Orin, $30-53K |
40-runtime-systems/data-logging/on-vehicle-data-triage-selective-upload.md | Vehicle-side data management: ring buffers (NVMe 1-4TB), event-triggered clips (safety/perception/localization), edge scenario classification, bandwidth-aware upload (50GB/day), compression, rosbag/mcap, fleet upload coordination, active learning integration |
60-safety-validation/runtime-assurance/online-perception-monitoring-odd-enforcement.md | ML silent degradation detection: input distribution monitoring, output consistency (CUSUM/EWMA), cross-modal agreement, OOD integration, ODD state machine with hysteresis, Perception Health Score, calibration drift, temporal anomaly, MLOps-scale runtime handoff, <5ms on Orin |
30-autonomy-stack/localization-mapping/maps/map-tile-versioning-distribution.md | Map distribution lifecycle: spatial tiling, content-addressable versioning (Merkle tree), differential updates (2-8% of full tile), atomic swap protocol, AIRAC integration, cryptographic signing, fleet synchronization, <500 MB/month |
50-cloud-fleet/fleet-management/ev-fleet-energy-co-optimization.md | Joint EV fleet energy co-optimization: EVRP formulation, LiFePO4 degradation, V2G demand response ($50-200/MWh), grid-aware scheduling, stochastic optimization, MILP/RL/MPC, OCPP 2.0.1 |
30-autonomy-stack/multi-agent-v2x/ramp-traffic-conflict-deadlock-prevention.md | Ramp traffic coordination: zone-capacity graph, reservation protocol, wait-die deadlock prevention, 9-level priority, stand sequencing, V2X fallback, MAPF (CBS/PIBT), $50-75K |
30-autonomy-stack/perception/overview/test-time-training-airport-onboarding.md | TTT for rapid airport onboarding: gradient-based auxiliary tasks, TTT-MAE, online LoRA, safety-bounded on Orin, catastrophic forgetting prevention, Simplex integration |
50-cloud-fleet/observability/fleet-anomaly-root-cause-attribution.md | Fleet anomaly attribution: CUSUM/EWMA monitoring, MLOps observability by scale, causal discovery (NOTEARS), Shapley values, Bayesian diagnosis, OTA regression, map staleness, environmental correlation, evidence artifacts |
50-cloud-fleet/data-platform/cloud-backend-infrastructure.md | Fleet data backend: three-zone data lake, S3+Lambda ingestion, MQTT streaming telemetry, Airflow orchestration, rosbag K8s processing, Feast feature store, auto-labeling, multi-airport isolation, $200-460/vehicle/month |
50-cloud-fleet/data-platform/data-catalog-lineage-quality-ops.md | Data catalog, lineage, and quality operations: data-product contracts, catalog/lakehouse/version-control choices, OpenLineage event design, quality gate matrix, SLOs, deletion propagation, promotion states, evidence artifacts, and release-blocking failure modes |
30-autonomy-stack/localization-mapping/maps/map-construction-pipeline.md | End-to-end HD map construction: survey drives → multi-session SLAM and LAMM-style map merging → alignment including RTK/GCP, OpenLiDARMap map-prior georeferencing, FlexCloud drift correction, and MapEval point-cloud map-quality QA → annotation → Lanelet2 → QA → deployment. 5-7 days at $20-35K per airport, AMDB bootstrap, SAM+CLIP auto-annotation |
20-av-platform/compute/edge-cloud-hybrid-inference.md | Three-tier compute (vehicle+edge+cloud): model placement, split inference, graceful degradation, Triton edge server, $2,500/vehicle shared edge, airport private 5G advantage |
20-av-platform/sensors/sensor-to-algorithm-readiness-contract.md | Sensor readiness contract: acquisition timestamps, calibration package, frame tree, preprocessing, health state, provenance, modality-specific checks, algorithm handoff table, reject/degrade rules, and evidence artifacts |
20-av-platform/sensors/calibration-bay-fixtures.md | Calibration bay fixtures: surveyed bay frame, target IDs, turntable/vehicle datum, LWIR/radar/LiDAR/camera evidence capture, residual gates, and signed calibration package handoff |
20-av-platform/sensors/automated-sensor-cleaning.md | Physical self-maintenance: air curtains + burst + washer + wiper + heated windows, contamination mapping, germanium-safe thermal cleaning, health monitor closed-loop, $200-500/vehicle, 15-25% availability gain |
20-av-platform/sensors/solid-state-lidar-photonics.md | Solid-state LiDAR: FMCW per-point velocity, silicon photonics OPA, Voyant Helium/Aeva Atlas/SiLC comparison, 100K+ hr MTBF, 1550nm eye safety, $150-450K/year fleet savings, phased migration strategy |
20-av-platform/networking-connectivity/deterministic-networking-tsn.md | Deterministic networking: IEEE 802.1 TSN (gPTP <100ns, TAS scheduling, FRER redundancy), safety messages <10μs (50-200x faster than CAN), mixed-criticality scheduling, CAN-TSN gateway, 5G TSN bridge, $230-440/vehicle |
30-autonomy-stack/vla-vlm/spatial-foundation-models-airport.md | Spatial foundation models: 4M, SpatialVLM, RT-2/Octo/pi0 for airport robotics, precision docking, FOD characterization, two-tier cloud+edge deployment, distillation for Orin |
50-cloud-fleet/fleet-management/fleet-predictive-maintenance.md | Fleet predictive maintenance: PHM framework, Weibull failure models, correlated airside failures, ML prediction, multi-echelon spare parts, cold-start sizing, joint scheduling, fleet availability modeling, 30-40% cost reduction |
30-autonomy-stack/planning/imitation-learning-behavioral-cloning.md | IL for airside: BC from teleop (BEV+GRU), MDN multimodal, Diffusion BC (DDIM 3-5 steps, 15-30ms Orin), DAgger with Frenet expert, MaxEnt IRL cost learning, GAIL, style-conditioned multi-operator, CBF post-processing, Simplex three integration modes, $35-55K |
30-autonomy-stack/planning/joint-prediction-planning.md | Joint prediction-planning: PDM-Closed baseline, conditional prediction, game-theoretic (Stackelberg, level-K), contingency planning, occupancy flow scoring, NAVSIM/nuPlan, SparseDriveV2/DiffusionDriveV2 metric caveats, Frenet augmentation with prediction costs (70-80% benefit at 10% cost) |
60-safety-validation/runtime-assurance/fail-operational-architecture.md | Fail-operational HW redundancy: 1oo2D, TMR, monitor-actuator, dual-Orin + FSI (DCLS R52 ASIL D), ASIL decomposition (ASIL B(D) + ASIL B(D)), sensor/actuator/power/CAN redundancy, degradation tiers (T0-T5), MRC planning, runway incursion HW geofence, $155-260K phased |
40-runtime-systems/ml-deployment/av-cicd-devops-pipeline.md | AV CI/CD pipeline: code CI (MISRA/static analysis), ML model CI (DVC/TensorRT), SIL/HIL/VIL simulation gates, map/config CI, artifact versioning, fleet deployment (canary rollout), ML regression detection, safety assurance, airside-specific requirements, and MLOps pipeline architecture by S0-S5 scale |
20-av-platform/compute/training-infrastructure.md | Training infrastructure and MLOps: cloud/on-prem GPU options, training cost estimates, MLOps scale controls, GPU FinOps, secure training metadata, and pipeline orchestration from manual research scripts to multi-tenant policy-gated training/evaluation platforms |
30-autonomy-stack/multi-agent-v2x/fleet-task-allocation-scheduling.md | Fleet GSE scheduling: MILP/CP-SAT optimal, CBBA decentralized, A-CDM predictive, RL dispatch, charging-aware, multi-objective |
60-safety-validation/runtime-assurance/weather-adaptive-odd-management.md | 5-level ODD with METAR/TAF/sensor fusion, capability curves, continuous speed envelope, jet blast zones, seasonal profiles |
30-autonomy-stack/localization-mapping/overview/robust-state-estimation-multi-sensor.md | ESKF deep dive, chi-squared gating, multi-hypothesis IMM, GPS-denied budgets, fleet state consistency, <0.5ms on Orin |
30-autonomy-stack/localization-mapping/maps/realtime-occupancy-grid-mapping.md | Log-odds occupancy, GPU raycasting, multi-LiDAR fusion, nvblox/VDBFusion, TSDF/ESDF, fleet-shared grids, costmap for Frenet |
50-cloud-fleet/data-platform/fleet-data-pipeline.md | End-to-end fleet data: 200GB/day/vehicle, DVC versioning, rosbag processing, MLOps-scale orchestration states, labeling workflows ($15-45/frame), fleet telemetry (Grafana), storage tiers, 5-100 vehicle scaling |
30-autonomy-stack/simulation/sim-to-real-transfer-airside.md | Sim-to-real for airside: LiDAR simulation fidelity, domain randomization, UniSim/LidarDM, curriculum learning, reality gap measurement, CARLA/Isaac airport env, $50-75K first airport |
30-autonomy-stack/perception/overview/test-time-adaptation-airside.md | TTA/domain adaptation for multi-airport: TENT, CoTTA, SAR, SFDA (SHOT/NRC), OOD triggers, active learning, LiDAR-specific adaptation, fleet-scale strategy, per-airport cost |
30-autonomy-stack/perception/overview/lidar-semantic-segmentation.md | LiDAR segmentation SOTA: Cylinder3D, FlatFormer, PTv3, SalsaNext; ALPINE training-free panoptic; 18-class airside taxonomy; Orin real-time (18-35ms); PointLoRA fine-tuning path |
30-autonomy-stack/perception/overview/aggregated-map-semantic-segmentation.md | End-to-end semantic segmentation of registered multi-scan LiDAR maps: aggregated-vs-single-scan, pipeline architecture, ML-related SLAM substrate scope, tiling/stitching, input modalities (LiDAR/RGB/fusion), compact proxy/input/training selector, source-map geometry QA through MapEval, large-scale datasets (Semantic3D, Paris-Lille-3D, Toronto-3D, DALES, SensatUrban), KPConv/RandLA-Net/sparse-conv/PTv3/Superpoint Transformer/point-cloud SSM backbones, permanence decision layer for map eligibility, LOSC pseudo-label consolidation, map-hygiene ground-truth gates, industry auto-label flywheel, airside L3 semantic layer |
30-autonomy-stack/perception/overview/model-compression-edge-deployment.md | Unified compression guide: PTQ/QAT quantization, knowledge distillation (TinyBEV), structured pruning, ModelOpt, per-model Orin recipes, 5-15x speedup at 1-3% accuracy loss |
30-autonomy-stack/perception/overview/multi-object-tracking.md | 3D MOT for airside: CenterPoint tracker, SimpleTrack, MCTrack, HOTA metrics, airside Re-ID (tail numbers, fleet IDs), ROS integration, 10Hz on Orin |
30-autonomy-stack/world-models/occupancy-deployment-orin.md | Occupancy on Orin: FlashOcc TensorRT (197 FPS), SparseOcc, LiDAR voxelization, nvblox ROS bridge, multi-resolution strategy, INT8 calibration |
20-av-platform/sensors/thermal-ir-cameras.md | Thermal cameras for airside: LWIR/MWIR bands, FLIR Boson 640 vs Seek Mosaic, night personnel detection, jet blast visualization, Orin MIPI integration, $8-22K/vehicle |
30-autonomy-stack/planning/neural-motion-planning.md | Neural/learned motion planning SOTA (2023-2026): IL planners (PlanTF, UniAD, VAD, SparseDrive/SparseDriveV2, GenAD, DiffusionDriveV2, Diffusion-Planner), game-theoretic (GameFormer, MARC), differentiable optimization (DIPP, DTPP), VLA planning (DriveVLM, Alpamayo, PlanAgent), safety (CBF, RSS, SafeDreamer, Simplex), NAVSIM PDMS/EPDMS benchmark caveats, Orin deployment |
30-autonomy-stack/localization-mapping/maps/hd-map-standards-airside.md | OpenDRIVE, AMDB/AMXM, NDS comparison; AMXM→Lanelet2 pipeline; NOTAM integration; cost estimates |
30-autonomy-stack/localization-mapping/maps/neural-online-mapping-sota.md | MapTracker (+69% consistency), StreamMapNet, NMP, topology reasoning (TopoMLP, LaneSegNet), airside adaptation strategy |
30-autonomy-stack/perception/overview/infrastructure-cooperative-perception.md | V2I fusion for airports, V2X-ViT/Where2comm/CoBEVT, QuantV2X/SparseCoop/VOGS-CP communication primitives, TruckV2X truck-centered proxy, existing airport systems (SMR/MLAT/ADS-B/CCTV), 0.5-1.5yr payback |
30-autonomy-stack/perception/overview/lidar-foundation-models.md | PTv3/Sonata/ScaLR, pre-training saves 50-80% labels, FlatFormer real-time on Orin, PointLoRA for fine-tuning |
30-autonomy-stack/localization-mapping/overview/lidar-slam-algorithms.md | KISS-ICP, LIO-SAM, FAST-LIO2, Point-LIO comparison; degeneracy detection; airside algorithm selection |
30-autonomy-stack/localization-mapping/overview/ml-related-slam-research-scope.md | ML-related SLAM research scope for semantic aggregated maps: learned registration, learned loop retrieval, semantic/dynamic SLAM, neural implicit and Gaussian maps, dynamic residual removal, motion/permanence/map-eligibility separation, static-but-transient quarantine, map priors, multi-session map merging, validation evidence, and segmentation handoffs |
60-safety-validation/cybersecurity/cybersecurity-airside-av.md | Threat models, ISO/SAE 21434, EASA requirements, sensor security, incident response |
70-operations-domains/deployment-playbooks/workforce-transition.md | 1.5-2M workers affected, union considerations, retraining, SATS case study |
90-synthesis/decisions/decision-framework.md | Architectural decision framework and diffusion planning guide |
30-autonomy-stack/vla-vlm/vlm-scene-understanding.md | VLM as co-pilot (not controller): DriveVLM CoT reasoning, DriveLM graph QA, NOTAM interpretation, turnaround status, FOD classification, anomaly detection, InternVL2-2B on Orin (300ms), $30-55K phased deployment |
60-safety-validation/verification-validation/airside-scenario-taxonomy.md | ISO 34502 adapted for airside ODD, Pegasus 6-layer model, 115 functional / 566 logical scenarios, SOTIF hazard catalog (H1-H8+), STPA control structure, risk matrix, testing strategy, regulatory coverage mapping |
70-operations-domains/airside/operations/ground-control-instructions.md | Airside instruction hierarchy (A-CDM→ATC→marshaller), A-SMGCS integration, D-TAXI digital clearance, NOTAM machine-readable parsing pipeline, marshaller gesture recognition (ViTPose+LSTM), NLU for ground control phraseology, instruction-to-trajectory mapping, conflict resolution priority, phased deployment $30-50K→$50-100K |
30-autonomy-stack/perception/overview/camera-fallback-perception.md | Camera-only degraded mode when LiDAR fails: Metric3D v2, DepthAnything v2 (15ms INT8 Orin), stereo depth (RAFT-Stereo, ZED 2i), BEVFormer-Tiny (35-50ms), confidence calibration, thermal stress, degraded mode architecture with speed reduction, Simplex integration |
70-operations-domains/deployment-playbooks/multi-airport-adaptation.md | Multi-airport scaling playbook: domain shift analysis, AMDB map bootstrapping (free FAA data saves 60-70% mapping cost), PointLoRA perception adaptation (500 labels), GNSS multipath mapping, seasonal adaptation, 8-week onboarding protocol, cost model ($75-150K per additional airport) |
40-runtime-systems/monitoring-observability/hmi-operator-interface.md | HMI design for airside AV: ISO 3691-4 operator interface, monitoring dashboard (ROS + Foxglove/web), trust calibration, 4-mode control architecture, handoff procedures (2-5s budget), operator training (40-80h), incident reporting → active learning, external crew communication (LED/audio), $5-15K per station |
30-autonomy-stack/localization-mapping/maps/semantic-mapping-learned-priors.md | Semantic maps + learned priors: Neural Map Prior (NMP, +5.4 mAP, +8.2 at night), PriorDrive unified prior encoding, T2SG topology scene graphs, conformal prediction for map uncertainty, fleet-based incremental map updates, 7-layer semantic map architecture, multi-airport LoRA adapters |
30-autonomy-stack/planning/safety-critical-planning-cbf.md | Formal safety for neural planners: CBF math framework (ECBF, HOCBF, stochastic/robust), neural CBF synthesis + conformal calibration (CP-NCBF), CBF-QP filter (<1ms on Orin), HJ reachability (DeepReach), game-theoretic planning (GameFormer level-K, GIME, Stackelberg), multi-agent CBFs (GCBF+ 1024 agents), CBF-Simplex three-layer architecture, airside-specific CBFs (aircraft proximity, jet blast, personnel, geofence, runway incursion) |
30-autonomy-stack/world-models/lidar-native-world-models.md | LiDAR-native world models: Copilot4D (>65% Chamfer improvement at 1s, >50% at 3s), UnO (self-supervised occupancy), LidarDM (diffusion LiDAR generation), LiDARCrafter (language-guided 4D), 4D occupancy forecasting, point cloud prediction networks, AD-L-JEPA, self-supervised training for airside, Orin deployment caveats, safety applications |
30-autonomy-stack/perception/overview/collaborative-fleet-perception.md | V2V cooperative perception: OPV2V/V2X-ViT/CoBEVT/CoBEVFlow SOTA, Where2comm bandwidth selection (95% perf at 1/64 bandwidth), HEAL heterogeneous agents, QuantV2X codebook-message quantization, SparseCoop sparse-query cooperation, VOGS-CP collaborative Gaussian occupancy, TruckV2X heavy-vehicle proxy, fleet occupancy map, collective FOD detection, 5G deployment, phased $15K→$115K |
30-autonomy-stack/planning/neuro-symbolic-scene-graphs.md | Neuro-symbolic reasoning: driving scene graphs, GNN interaction modeling (LaneGCN, HiVT, HDGT), knowledge graphs for traffic rules, STL-constrained planning (differentiable), compositional reasoning, LLM-symbolic hybrid, airport right-of-way encoding (9-level priority), NOTAM rule injection, interpretable decisions, certification argument structure |
60-safety-validation/verification-validation/testing-validation-methodology.md | AV testing methodology: V-model, ASAM OpenSCENARIO 2.0, N-wise covering arrays (1,280→40 tests), CMA-ES falsification, LLM scenario generation, metamorphic testing, SIL/HIL/VIL, Zhao-Weng formula (4,600 tests for 99.9% reliability), Bayesian safety, shadow mode criteria, regression CI/CD, digital twin, $105K first airport |
30-autonomy-stack/perception/overview/self-supervised-pretraining-driving.md | Unified SSL pre-training: contrastive (SLidR, ScaLR, PPKT), MAE (Voxel-MAE, GD-MAE, BEV-MAE), JEPA (AD-L-JEPA, V-JEPA 2), DINOv2 for driving, multi-modal pre-training (UniPAD, BEVDistill), LoRA fine-tuning, 50-80% label reduction, airside curriculum (road SSL→road supervised→airside SSL→airside supervised), $5-15K compute vs $80K+ labeling |
30-autonomy-stack/perception/overview/gaussian-splatting-driving.md | 3DGS for real-time perception/mapping: GaussianFormer (39.2 mIoU, 20 FPS, 3.2x less memory), GaussianFormer v2 (41.1 mIoU), VOGS-CP collaborative Gaussian occupancy, GaussianOcc self-supervised (80% gap closure, zero labels), SplaTAM SLAM (<0.4cm ATE), MonoGS, LiDAR-Gaussian fusion, multi-LiDAR merging via covariance intersection, dynamic object tracking, semantic/panoptic Gaussians, LangSplat language grounding, FOD detection via map anomaly, aircraft proximity monitoring, hybrid PointPillars+GaussianFormer architecture, Orin ~92ms, $90K/12-18mo integration |
50-cloud-fleet/mlops/data-flywheel-airside.md | Closed-loop data flywheel: trigger-based collection (50GB/day/vehicle, 100% safety capture), auto-labeling (SAM+CLIP foundation models, 70-85% cost reduction to $1.50-3/frame), semantic-map pseudo-label promotion gated by release-state labels, active learning (40-50% fewer labels, safety-weighted), label budget and QA controls by MLOps scale, continuous retraining (monthly cycle), shadow mode validation (1-2 weeks), A/B fleet testing, scenario mining (power-law long-tail), synthetic data ($23K for 35K frames), multi-airport LoRA ($2-8K/airport), mAP trajectory 45%→82% over 24mo, breakeven Month 18, $205K Year 1 |
50-cloud-fleet/mlops/model-governance-release-evidence.md | Model governance and release evidence: registry aliases, claims-and-evidence release reviews, ownership matrix, offboard labeler/prompt evidence, incident and rollback evidence by S0-S5 MLOps scale, executable rollback requirements, and release failure modes |
50-cloud-fleet/mlops/mlops-scale-research-scope.md | MLOps scale research scope: maturity ladder from notebook research and repeatable prototypes to production product, fleet/multi-site autonomy, regulated safety-critical release, and foundation-model/platform scale; covers data contracts, lineage, split/leakage firewalls, label ops, experiment tracking, orchestration, compute, registries, evaluation, replay, shadow/canary release evidence, serving, monitoring, governance, reference-architecture routing, build order, transition triggers, and autonomy-specific map/model/calibration evidence |
50-cloud-fleet/mlops/mlops-reference-architectures-by-scale.md | MLOps reference architectures by scale: S0-S5 blueprints, minimum viable stack, centralize/decentralize/delay decisions, durable manifest interfaces, product/fleet/regulatory/platform lanes, migration sequence, airside managed-site target, and platform failure modes |
50-cloud-fleet/mlops/mlops-migration-checklist-by-scale.md | MLOps migration checklist by scale: artifact-authority migration principles, S0-S5 entry/exit criteria, transition checklists, workstream migration matrix, tooling upgrade triggers, 30/60/90 rollout plan, managed-site caveats, migration evidence packet, and failure modes |
50-cloud-fleet/mlops/mlops-scorecards-and-kpis-by-scale.md | MLOps scorecards and KPIs by scale: reproducibility, data lineage, label quality, model/runtime quality, release reliability, observability, incident response, governance, cost, release blockers, cadence, ownership, managed-site KPI focus, and anti-metrics |
50-cloud-fleet/mlops/experiment-tracking-reproducibility-by-scale.md | Experiment tracking and reproducibility by scale: scratch/exploratory/baseline/candidate/release/evidence/platform-benchmark authority states, R0-R5 reproducibility levels, run manifest contract, MLflow/W&B/DVC/TensorBoard/MLMD/OpenLineage architecture comparison, run comparison rules, LiDAR-image and ML-SLAM lineage, managed-site rules, scorecards, and failure modes |
50-cloud-fleet/mlops/model-registry-artifact-lifecycle-by-scale.md | Model registry and artifact lifecycle by scale: registry scope for model weights, runtime packages, semantic maps, calibration bundles, prompt/labeler/evaluator packs, replay/eval packs, feature snapshots, evidence bundles, alias policy, lifecycle states, artifact-set records, architecture comparison, deletion/retention, and managed-site semantic-map rules |
50-cloud-fleet/mlops/serving-inference-operations-by-scale.md | Serving and inference operations by scale: serving ownership, S0-S5 ladder, batch/online/shadow/canary/edge modes, architecture comparison for Triton, KServe, Seldon, Ray Serve, BentoML, MLServer, ONNX/TensorRT, and managed endpoints, service manifest fields, runtime optimization, traffic routing, autonomy managed-site rules, KPIs, and failure modes |
50-cloud-fleet/mlops/platform-sre-reliability-by-scale.md | MLOps platform SRE and reliability by scale: platform control-plane reliability, criticality tiers, SLIs/SLOs, error budgets, backup/restore, DR, managed cloud versus self-hosted patterns, incident response, tenant/site isolation, managed-site autonomy rules, release blockers, KPIs, and failure modes |
50-cloud-fleet/mlops/pipeline-orchestration-release-workflows-by-scale.md | Pipeline orchestration and release workflows by scale: workflow ownership, S0-S5 orchestration ladder, workflow authority states, orchestrator comparison across scripts/DVC/GitHub Actions/Airflow/Argo/Kubeflow/TFX/Ray/Slurm/managed platforms, durable pipeline interfaces, release state machine, autonomy workflow families, scheduler coupling, managed-site rules, scorecards, and failure modes |
50-cloud-fleet/mlops/evaluation-platform-replay-gates-by-scale.md | Evaluation platforms and replay gates by scale: evaluation ownership, S0-S5 authority states, evaluation manifest contract, architecture comparison across scripts, CI, MLflow Evaluate, TFMA, Evidently, managed cloud eval, custom replay services, simulation, and scorecards, with LiDAR/image semantic-map evaluation and training-architecture coupling |
50-cloud-fleet/mlops/dataset-split-leakage-controls-by-scale.md | Dataset split and leakage controls by scale: split manifests as release artifacts, leakage taxonomy, S0-S5 controls, manifest contract fields, training/evaluation architecture comparison, LiDAR/image/ML-SLAM rules, managed-site non-road holdouts, scorecards, and failure modes |
50-cloud-fleet/mlops/model-monitoring-drift-response-by-scale.md | Model monitoring and drift response by scale: monitoring signal taxonomy, S0-S5 response authority, architecture comparison, monitoring event contract, response state machine, trigger-to-action matrix, retraining trigger policy, managed-site rules, KPIs, and failure modes |
50-cloud-fleet/mlops/site-sliced-release-evidence-by-scale.md | Site-sliced release evidence by scale: ODD-cell release units, S0-S5 evidence ladder, release manifests, managed-site slice taxonomy, release state machine, training/adaptation tradeoffs, statistical discipline, semantic-map/ML-SLAM coupling, acceptance checks, and failure modes |
50-cloud-fleet/mlops/feature-embedding-store-ops-by-scale.md | Feature and embedding store operations by scale: manifest-backed files, offline feature stores, online feature stores, lakehouse/catalog tables, vector indices, autonomy use cases, contract fields, training architecture comparison, invalidation/backfill rules, monitoring, and failure modes |
50-cloud-fleet/mlops/offboard-labeler-registry-by-scale.md | Offboard labeler registry by scale: governed identity, state machine, output-state contract, prompt/model/retrieval/threshold records, evaluation gates, architecture comparison, semantic-map/non-road rules, scorecards, and failure modes for auto-labelers and foundation-model labelers |
50-cloud-fleet/mlops/gpu-queueing-finops-by-scale.md | GPU queueing and FinOps by scale: workload classes, S0-S5 capacity ladder, release/evidence/incident priority queues, scheduler comparison, unit economics, job metadata contract, capacity planning, safety-evidence policy, monitoring scorecards, and failure modes |
50-cloud-fleet/mlops/secure-artifact-attestation-profile.md | Secure artifact attestation by scale: artifact scope, S0-S5 control profile, attestation types, trust-chain architecture, registry alias policy, policy-enforcement options, verification gates, managed-site mapping notes, metadata schema, and failure modes for signed MLOps release artifacts |
50-cloud-fleet/mlops/federated-privacy-preserving-training-policy-by-scale.md | Federated and privacy-preserving training policy by scale: decision frame, trigger scorecard, architecture comparison, required client/update/privacy contracts, privacy controls, training architecture comparison, evaluation gates, managed-site rules, operating model, and failure modes |
50-cloud-fleet/mlops/llmops-agent-evaluation-by-scale.md | LLMOps and agent evaluation by scale: artifact scope, S0-S5 policy, autonomy use cases, architecture comparison, evaluation layers, agent trajectory contract, release gates, observability, managed-site rules, and failure modes |
50-cloud-fleet/mlops/map-derived-pseudo-label-invalidation-protocol.md | Map-derived pseudo-label invalidation protocol: trigger taxonomy, active/suspect/quarantined/rebuilt/reapproved/deprecated states, lineage impact graph, batch manifest fields, scale controls, and airside release-state rules |
50-cloud-fleet/ota/perception-slam-artifact-compatibility-matrix.md | Perception-SLAM artifact compatibility matrix: MLOps-scale compatibility posture, signed artifact-set manifest, model/map/calibration/runtime/telemetry/semantic-taxonomy/prompt-labeler/replay compatibility, release gates, rollback, quarantine, and SUMS governance |
50-cloud-fleet/data-platform/replay-scenario-mining-ops.md | Replay and scenario mining operations: candidate-to-regression scenario state machine, ASAM OpenSCENARIO/OpenLABEL-aligned artifacts, semantic-label evidence, clean-worker replay packages, and MLOps scale guidance from lightweight S0 clips to S5 shared scenario catalogs with coverage, flake, age, duplicate, and cost controls |
50-cloud-fleet/operations/fleet-sre-incident-response.md | Fleet SRE and incident response: AV severity taxonomy, incident command roles, evidence manifests, fleet/site/ODD containment, MLOps incident response by scale, blast-radius queries across model/map/calibration/runtime/prompt/data artifacts, and safety-case deltas |
50-cloud-fleet/ota/software-update-management-system-ops.md | SUMS operations: update classification, impact analysis, validation bundles, risk-based rollout, post-deployment closure, rollback controls by MLOps scale, emergency-update evidence completion, and software/model/map/config/calibration governance |
60-safety-validation/runtime-assurance/runtime-verification-monitoring.md | Runtime verification: STL quantitative robustness as unified safety metric, 20 airside-specific STL specs (aircraft proximity, zone speed, geofence, runway incursion, jet blast), RTAMT tool for ROS, combined OOD detection (energy+Mahalanobis+ensemble, 95-98% AUROC), conformal prediction coverage guarantees, 9 airside OOD triggers, maximally permissive shields (1-5% intervention), Shield+CBF+Simplex three-layer defense-in-depth, safety MCU (STM32H725, $50-200/vehicle), METAR→ODD monitoring, WCET <5.5ms full suite, ISO 26262 ASIL decomposition, UL 4600 compliance, DO-178C formal methods credit, fleet-level anomaly correlation, $115-200K/32 weeks |
30-autonomy-stack/world-models/occupancy-flow-4d-scenes.md | Occupancy flow & 4D scene understanding: static→dynamic occupancy, scene flow (NSFP, ZeroFlow 0.028m EPE3D, DeFlow 0.023m SOTA), 4D forecasting (UnO self-supervised winner, OccSora diffusion, Cam4DOcc benchmark, SelfOccFlow), dynamic 3DGS (StreetGaussians, 4D-GS, K-Planes 10900x compression), flow-guided Frenet planning (60-70% collision reduction), temporal modeling (attention+GRU hybrid), sparse voxels (18x compression), Orin 26-40ms FP16 pipeline, class-agnostic motion prediction, $6-11K training cost |
30-autonomy-stack/perception/overview/streaming-temporal-perception.md | Streaming temporal perception: StreamPETR object-centric propagation (+6-8% NDS, <3ms overhead, implicit AMOTA 65.3%), Sparse4D v3 (71.9% NDS SOTA), multi-sweep LiDAR (3-sweep +2.5% mAP at +1.4ms), BEV temporal fusion (BEVFormer +10.1% NDS), latency-aware streaming (ASAP/LASP), temporal consistency filtering (eliminates de-icing/jet blast transients), extended airside track persistence (30s GSE, 300 frames aircraft), turnaround phase detection, video backbone comparison, $38K/13 weeks |
30-autonomy-stack/perception/overview/active-perception-sensor-scheduling.md | Active perception & sensor scheduling: context-aware model switching (35-45% compute savings), entropy-based attention allocation, foveated LiDAR (89% voxel reduction), multi-LiDAR scheduling (3-4 of 8 at full, 44% savings), early exit networks (48% average compute), risk-aware allocation, planner-guided attention, predictive load scheduling via A-CDM, safe model switching (3-frame overlap), 30-36% power savings for electric GSE, $25-40K/10 weeks |
60-safety-validation/verification-validation/formal-verification-neural-networks.md | Formal verification of neural networks: SMT (Reluplex, Marabou) and MILP for complete verification (<100K params), alpha-beta-CROWN over-approximation (VNN-COMP winner, millions of params), DeepPoly/PRIMA abstract interpretation, IBP/SABR certified training, Lipschitz bounds for safety margins, randomized smoothing, layered strategy (complete for policy/CBF/Simplex, scalable for PointPillars/CenterPoint, runtime for residual), auto_LiRPA code examples, ISO 3691-4/UL 4600/EU AI Act/EU Machinery Regulation compliance |
20-av-platform/compute/energy-efficient-inference-24-7.md | Energy-efficient 24/7 inference: Orin 15W/30W/50W power modes deep dive, dynamic model switching (40-60% low-complexity time), thermal management (-10C to +50C tarmac, throttling curves), battery-aware compute (SoC-correlated budgets), DLA offloading (5-10W concurrent), sleep/wake (<500ms wake-up), per-model watt profiling, fleet-level energy optimization, 8-15% more daily operating hours, 12-18C lower junction temperature, $15-25K implementation |
30-autonomy-stack/planning/reinforcement-learning-driving-policy.md | RL driving policy: CaRL (CoRL 2025 SOTA, PPO + route completion reward scales with batch size), IQL (best offline RL, consistent across traffic densities), SAC/TD3/TQC/CrossQ off-policy comparison, BC→offline RL→online RL three-phase pipeline, CQL conservative lower-bound Q-values, Decision Transformer (RL as sequence modeling), safe RL (CPO, Lagrangian PPO, CBF-QP filter decouples safety from performance), Recovery RL (emergency maneuvers), privileged-to-sensor distillation (comma.ai approach), DAgger with Frenet planner as oracle, RLPD 50/50 mixing for offline-to-online, policy head 0.5ms FP16 on Orin, Simplex integration (RL advanced + Frenet fallback), $45-75K over 32 weeks |
30-autonomy-stack/localization-mapping/maps/hd-map-change-detection-maintenance.md | HD map change detection and maintenance: point cloud differencing (ICP-based, KD-tree), semantic change detection (class-based filtering), RTMap (ICCV 2025, centimeter-level recursive map maintenance), Bayesian fleet consensus (per-vehicle reliability, posterior >0.99 for safety-critical), DBSCAN spatial clustering, temporal decay model (feature-type half-lives: structures 365d, barriers 30d, equipment 7d), AIRAC 28-day cycle integration (dual-layer: regulatory AIRAC + operational fleet), light-map alternative (720 KB topology+safety+regulatory), NMP implicit maintenance, 3DGS map updates (opacity decay), OTA canary deployment (10% fleet first, 2h monitoring), construction zone + NOTAM corroboration, cost: $45-70K/28 weeks, 60-80% reduction vs manual re-survey, break-even at 2-3 airports |
70-operations-domains/airside/business-case/fleet-tco-business-case.md | Fleet TCO and business case: per-vehicle CAPEX ($95-210K floor at scale), LiDAR-only sensor kit $29-60K, full suite $47-84K, vehicle integration $20-30K, 3-shift labor savings $150K/year per position, accident avoidance $150-750K/year for 20 vehicles, scale dynamics (pilot $400-650K/vehicle → mature $155-330K), multi-airport marginal cost $600K→$115K, certification $530K-1.95M across 5 jurisdictions, operator ratio 1:5→1:10+ as key OPEX lever, break-even Year 2-4, 10-year NPV $45-80M at 200 vehicles (8% discount), RaaS $10-14K/month, probability-weighted expected NPV ~$25M, regulatory delay -$8-15M/year NPV impact, UISEE 40-60% cost advantage, airport cluster deployment strategy, $2-6B TAM at 10% penetration |
30-autonomy-stack/multi-agent-v2x/v2x-protocols-airside.md | V2X communication protocols for airside: C-V2X over private 5G/CBRS (preferred, sub-ms URLLC), DSRC comparison, ETSI ITS message architecture (CAM 1-10 Hz, DENM events, CPM perception sharing, MCM maneuver coordination), 8 custom airside messages (Aircraft Proximity Alert, Stand Operation Status, GSE Task Assignment, De-Icing Zone, Emergency Vehicle Priority, Runway Incursion Prevention default-deny model, FOD Detection Alert, Jet Blast Warning — highest criticality invisible hazard), protobuf field-level specs with example payloads, A-CDM/A-SMGCS/ADS-B/AODB bridge architecture, bandwidth planning (123 Mbps for 50 vehicles, zone filtering needed at 200+), PKI with airport-managed CA hierarchy, misbehavior detection trust scoring, fallback safe behavior without V2X (5 km/h + 2x margins), cooperative perception +15-25% AP, standards predicted 2028-2030, $270-450K full implementation, V2X hardware $200-600/vehicle on existing 5G |
Document Statistics
| Metric | Value |
|---|---|
| Reader Markdown pages | 865 |
| Core research documents | 861 |
| Reader/research lines | 390k+ |
00-start-here/ documents | 4 |
10-knowledge-base/ documents | 137 |
20-av-platform/ documents | 41 |
30-autonomy-stack/ documents | 458 |
40-runtime-systems/ documents | 23 |
50-cloud-fleet/ documents | 41 |
60-safety-validation/ documents | 56 |
70-operations-domains/ documents | 27 |
80-industry-intel/ documents | 62 |
90-synthesis/ documents | 12 |
| Companies covered | 25 |
| Technology domains | 9 |
| Method-level SLAM library | 158 SLAM-method documents including overview/audit |
| Method-level perception files | 138 |
| Safety and validation documents | 56 |
| AV platform documents | 41 |
| Knowledge base documents | 137 |
| Synthesis documents | 12 |
| Perception documents | 212 |
| Localization/mapping | 182 |
| Planning documents | 16 |
| Multi-agent and V2X | 7 |
| Robustness validation files | 9 |
| Papers referenced | 700+ |
| Open-source repos evaluated | 90+ |
| Occupancy methods compared | 20 |
| Online mapping methods compared | 16 |
| Cooperative perception methods | 10+ |
| Airport deployments documented | 15+ |