Continuous Research Loop
This page turns the repo's gap audits into a repeatable research loop. It exists so perception, SLAM, mapping, sensors, and platform coverage keep moving from discovery to atomic files instead of remaining as broad backlog rows.
For repeatable source monitoring, use the Active Frontier Source Registry. It records the sites, filters, query patterns, review cadence, and automation feasibility for active frontier discovery before candidates are routed into the canonical audits or backlog.
Loop Contract
| Stage | Output | Done when |
|---|---|---|
| Discover | Web-search and repo-audit findings | Missing techniques, methods, datasets, fundamentals, and platform topics are named with primary sources. |
| Triage | P0/P1/P2 queue | Each item has a target directory, owner type, and reason it matters to AV, indoor/outdoor, mapping, or airside use. |
| Promote | Atomic research files | One method, technique, sensor model, or platform primitive gets its own source-backed file. |
| Cross-link | Overviews, audits, README, INDEX, methodology | Readers can find the new file from the static portal without knowing the path. |
| Verify | Link checks, tests, build, stale-path scan | The corpus remains navigable and deployable through VitePress. |
| Repeat | Next queue | Remaining backlog items are smaller, better prioritized, and ready for the next agent wave. |
Current Promotion Waves
The 2026-05-09 loops promoted thirteen high-value gap clusters into first-class pages. The follow-up Perception and SLAM Gap Fill and Cross-Architecture Knowledge Base Gap Fill record the next source-backed queues discovered by the latest scout and research-agent passes; the cross-architecture follow-up also promoted 30 additional atomic pages.
| Track | Promoted coverage |
|---|---|
| Gaussian and 4D perception | SplatAD, GaussianFormer, GaussianOcc, streaming Gaussian occupancy, Cam4DOcc, StreamingFlow, Sparse4D, TacoDepth, and RaCFormer. |
| SLAM and mapping methods | MOLA, KISS-SLAM, KISS-Matcher, LVI-SAM, FAST-LIVO/FAST-LIVO2, R2LIVE/R3LIVE, Splat-SLAM, S3PO-GS, Gaussian-LIC, GS-LIVM, VIGS-SLAM, dynamic 4D Gaussian SLAM, and RadarSplat-RIO. |
| Sensor and estimation fundamentals | LiDAR, camera, IMU, GNSS/RTK, radar, event/thermal, time synchronization, multi-sensor calibration observability, wheel odometry, visible-camera hardware, and IMU/GNSS/RTK hardware. |
| First-principles foundations | Gaussian noise, Mahalanobis gating, MAP/MLE, robust statistics, mixtures, Gauss-Newton, LM, dogleg, Cholesky, QR/SVD, sparse solvers, Lie groups, PnP, ICP/GICP/NDT, occupancy grids, data association, JPDA/MHT/RFS, filters, sensor likelihoods, signal processing, radar ambiguity, CFAR, timestamping, and statistical benchmarking. |
| LiDAR artifact removal and map cleaning | LIORNet, LiSnowNet, SLiDE, TripleMixer, 3D-KNN blind-spot de-snowing, 3D-OutDet, AdverseNet, DenoiseCP-Net, classical outlier filters, broad weather artifact removal, LiDAR ghost/multipath artifacts, ERASOR, Removert, dynamic map cleaning, and artifact-removal validation. |
| ML foundations for autonomy | Perceptrons, logistic/softmax cross-entropy, MLPs, backprop/autodiff, optimization dynamics, initialization/normalization/regularization, CNNs, RNN/LSTM/GRU, attention/transformers, vision transformers, SSL, sequence models, foundation training, JEPA, and world-model first principles. |
| Dynamic/static object removal | MapCleaner, ERASOR++, 4dNDF, FreeDOM, STATIC-LIO dynamic-point removal, MotionSeg3D, MambaMOS, neural scene-flow priors, moving/static separation datasets, moved-object map-change datasets, scene-flow benchmarks, 4D occupancy benchmarks, and airside dynamic-map cleaning validation. |
| ML objective and evaluation foundations | Autoencoders/VAEs, contrastive InfoNCE, masked modeling, EBMs, tokenization/discretization, positional encodings, S4/Mamba first principles, diffusion-score-flow samplers, multi-task losses, calibration/leakage, and world-model evaluation objectives. |
| Latest perception/radar/neural-field gaps | CVFusion, 4D radar-camera occupancy, POD/FMCW LiDAR predictive detection, DrivingGaussian, HUGS, SplatFlow, DistillNeRF, TrackOcc, cross-domain LiDAR scene flow, self-supervised occupancy flow, and UniScene occupancy-centric generation. |
| Latest SLAM robust/lifelong gaps | Robust PGO/GNC/riSAM, certifiable pose-graph optimization, Kimera-RPGO/PCM, distributed multi-robot PGO, LT-mapper/Khronos, lifelong 3D map version control, detector-based potentially dynamic object removal, RTMap/DUFOMap, GPR localization, radar teach-repeat, and MOVES. |
| KB probability/control foundations | Probabilistic graphical models and message passing, information theory for perception/ML, uncertainty quantification/calibration/conformal prediction, constrained optimization/MPC/iLQR, and MDP/POMDP/belief-space/RL foundations. |
| Dataset and validation gaps | MUSES, sensor-corruption robustness benchmarks, open-world/OOD anomaly segmentation benchmarks, FOD and airport apron detection datasets, FOD perception validation, and knowledge-base evaluation protocol. |
| Perception/SLAM reliability gap wave | LiDAR-camera occupancy fusion, dynamic occupancy/free-space, 4D radar road-boundary/freespace detection, radar-LiDAR adverse-weather detection, RobuRCDet, SAMFusion, STU, synthetic multimodal FOD benchmarks, 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/4D-radar localization, continuous-time and volumetric-map foundations, detection/tracking foundations, fleet-data contracts, and perception/SLAM/map validation protocols. |
| Cross-architecture KB gap fill | 30 pages across safety standards, runtime/data governance, platform integration, closed-loop autonomy evaluation, regulatory/deployment maps, and company deployment evidence. |
| MLOps scale architecture loop | Evaluation/replay gates, registry lifecycle, serving/inference operations, platform SRE/reliability, and data catalog/lineage/quality operations by scale now cover S0-S5 release evidence, artifact identity, data-product contracts, lakehouse/catalog/version-control choices, OpenLineage-style event boundaries, quality gate severity, endpoint/batch/edge serving manifests, service criticality tiers, SLOs, error budgets, backup/restore, DR, tenant isolation, deletion propagation, traffic routing, autoscaling, ODD-cell canary, rollback, monitoring, governance, and OTA compatibility. |
| Sparse-query and end-to-end driving (2026-05-23) | SparseBEV, SparseDrive, and DiffusionDrive promoted into the perception method library. |
| Recurring perception occupancy loop | DepthOcc as a depth-aware, real-time camera occupancy method with Occ3D-nuScenes evidence and explicit non-road transfer notes. |
| Recurring SLAM benchmark loop | COSMO-Bench as a collaborative SLAM optimization benchmark with multi-robot graph data, communication variants, and loop-outlier labels. |
| Recurring sparse-query perception loop | SparseBEV and DETR4D as atomic sparse-query multi-camera 3D detection pages, with DriveBench kept as an existing VLA/VLM reliability cross-link. |
| Recurring occupancy follow-on loop | TEOcc as a radar-camera temporal occupancy method and GaussRender as a Gaussian-rendered projective supervision module for 3D occupancy. |
| Recurring adverse-weather dataset loop | DSERT-RoLL and CMHT as multimodal public proxies for event/thermal/radar/LiDAR adverse-weather fusion, plus an explicit no-public-data gap for airside dust, de-icing mist, steam, glycol film, and wet-apron multipath. |
| Recurring practical SLAM loop | SLAM Toolbox as the ROS 2/Nav2 indoor 2D SLAM reference, NDT variants and maps as the split-out page for NDT-D2D, NDT-OM, NDT-MCL, multi-resolution maps, and dynamic map loading. |
| Recurring visual/VIO benchmark loop | LaMAria, Hilti x Trimble SLAM Challenge 2026, and ScaleMaster as current visual-inertial, floor-plan localization, and monocular scale/map-quality benchmark pages. |
| Recurring Gaussian/foundation SLAM loop | HI-SLAM2, SEGS-SLAM, and a neural/Gaussian SLAM survey page as current visual 3DGS SLAM method and taxonomy coverage. |
| Recurring radar/SLAM hardening loop | GV-iRIOM as a globally referenced 4D radar/visual/GNSS mapping page, plus a refreshed radar-inertial online calibration page covering LC-RIO-ET-style spatio-temporal calibration. |
| Recurring world-model maintenance loop | Drive-OccWorld source IDs and AAAI year corrected, occupancy-world-model action-conditioning wording tightened, and DFIT-OccWorld/IR-WM/SparseWorld/OccTENS/OccSim follow-ons routed into existing world-model overview pages. |
| Recurring Gaussian occupancy loop | GaussianFlowOcc and GaussTR as source-mature Gaussian occupancy method pages, plus LinkOcc routing/source correction so RobuRCDet is not cited as a LinkOcc source. |
| Recurring GS-Occ3D occupancy loop | GS-Occ3D as a vision-only Gaussian-surfel occupancy reconstruction and binary-label curation method page; LinkOcc remains watchlist-only until open method details or code make a page reviewable. |
| Recurring SLAM benchmark matrix refresh | Oxford Spires, IILABS 3D, SMapper-light, FusionPortableV2, and S3E were routed into the SLAM benchmark matrix with source caveats instead of creating duplicate dataset pages. |
| Recurring audit/navigation cleanup | Count drift and duplicate-prone queue wording were cleaned so future loops do not re-promote EvOcc, DR-REMOVER, ExelMap, or already-existing perception pages. |
| Recurring world-model hygiene loop | Copilot4D action-conditioning and Chamfer wording corrected, Drive-OccWorld LiDAR-adaptation and planning-metric claims tightened, DiffusionDrive nuScenes values corrected, and NAVSIM PDMS/EPDMS rows split by metric and benchmark. |
| Recurring calibration-bay fixture loop | Calibration-bay fixtures promoted as a physical platform page covering surveyed bay frames, target inventories, turntable/vehicle datum, LWIR/radar/LiDAR/camera evidence capture, release gates, and airside transfer notes. |
| Recurring cooperative compression loop | QuantV2X promoted as a cooperative perception quantization page covering codebook-index BEV messages, full-stack PTQ, deployment latency claims, domain transfer caveats, and runtime artifact gates. |
| Recurring sparse cooperative-query loop | SparseCoop promoted as a cooperative sparse-query detection/tracking page covering kinematic-grounded queries, coarse-to-fine aggregation, cooperative instance denoising, V2X-Seq/Griffin metrics, and source-mature prototype caveats. |
| Recurring truck-centered V2X dataset loop | TruckV2X promoted as a truck-centered cooperative perception dataset page covering tractor/trailer/CAV/RSU agents, 64 synthetic scenarios, 88,396 LiDAR frames, 1.18M boxes, occlusion recovery analysis, and articulated-vehicle transfer caveats. |
| Recurring collaborative Gaussian occupancy loop | VOGS-CP promoted as a collaborative Gaussian semantic occupancy method covering sparse 3D semantic Gaussian messages, ROI-culling, cross-agent Gaussian fusion, Gaussian-to-voxel splatting, AAAI 2026 source maturity, and simulation-to-real transfer caveats. |
| Recurring FOD synthetic data-engine loop | Airport-FOD3S promoted as a synthetic FOD data-engine page covering real seed data, size/seam/style-controlled FOD insertion, detector ablations, synthetic sample lineage, real-only validation gates, DualFOD/FOD-UAS benchmark routing, and LIDAROC sensor-contamination proxy routing. |
| Recurring raw GNSS factor fusion loop | GVINS/GLIO promoted as a raw GNSS factor-fusion page covering pseudorange and Doppler factors, visual-inertial and LiDAR-inertial companions, urban-canyon caveats, antenna lever-arm and receiver-clock assumptions, and multipath-safe validation gates. |
| Recurring wheel/vehicle-motion factor loop | Wheel odometry and vehicle-motion factors 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, vehicle-mode assumptions, and validation gates. |
| Recurring embodied 3D benchmark loop | EmbodiedScan/MMScan 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 learned 4D radar odometry loop | CAO-RONet promoted as a learned radar-only odometry page covering local completion, context-aware association, clip-window optimization, View-of-Delft evaluation context, repository maturity, and all-weather transfer caveats. |
| Recurring Doppler radar bridge loop | Doppler Radar-LiDAR SLAM promoted as a bridge page for Radarize, DRO, and Doppler-SLAM, keeping Doppler-aware radar odometry, direct radar registration, and radar-LiDAR-inertial fusion routed without splitting duplicate one-method pages. |
| Recurring radar RIO correspondence/uncertainty loop | Radar RIO Correspondence and Uncertainty promoted as a combined hardening page for learned radar point correspondences, polar point uncertainty, continuous point-pose uncertainty, and learned velocity uncertainty routing. |
| Recurring radar place-recognition refresh loop | Radar Place Recognition refreshed as a 4D radar descriptor lineage page routing 4DRaL, SHeRLoc, TransLoc4D, 4D RadarPR, and TDFANet without creating duplicate one-method pages. |
| Recurring planning/E2E refresh loop | SparseDriveV2 and DiffusionDriveV2 routed into existing neural planning, diffusion planning, joint prediction-planning, E2E architecture, and NAVSIM/Bench2Drive benchmark pages with PDMS/EPDMS separation and simulation-to-real caveats. |
| Recurring open-vocabulary 3D segmentation loop | SpaCeFormer promoted as a fast proposal-free open-vocabulary 3D instance segmentation page with NVIDIA project, arXiv, and WarpConvNet source links; DySS remains routed through sparse-query and streaming-temporal overviews pending official code/adoption. |
| Recurring active calibration design loop | Active Calibration Experiment Design promoted as a first-principles geometry page covering Fisher-information observability metrics, active trajectory/data collection, targetless online calibration caveats, safe ODD-bounded excitation, and release evidence routing. |
| Recurring robust-loss covariance loop | Robust-Loss Covariance Consistency promoted as a state-estimation foundation page covering covariance reporting after robust weighting, gating, GNC, Ceres/GTSAM covariance choices, selection-bias diagnostics, and NIS/NEES release evidence. |
| Recurring two-view geometry loop | Epipolar Geometry, Homographies, and Two-View Verification promoted as a geometry foundation page covering fundamental and essential matrices, homography model choice, robust inlier masks, cheirality, planar/low-parallax degeneracy, rolling-shutter caveats, and AV release diagnostics. |
| Recurring optical/scene-flow foundations loop | Optical Flow and Scene Flow First Principles promoted as a geometry foundation page covering brightness constancy, Lucas-Kanade local least squares, Horn-Schunck smoothness, depth lifting, rigid/residual scene flow, point-cloud flow objectives, occupancy-flow handoff, domain transfer caveats, and release diagnostics. |
| Recurring constrained-solver mechanics loop | Constrained KKT, QP, and SQP Solver Mechanics promoted as an optimization foundation page covering KKT stationarity/feasibility/complementarity, QP forms, active-set/ADMM/interior-point/SQP solver families, MPC warm starts, infeasibility/slack handling, active-constraint telemetry, and release evidence. |
| Recurring AV data-evaluation foundations loop | AV Data Evaluation Fundamentals promoted as a machine-learning foundation page covering split manifests, scenario/ODD coverage, leakage firewalls, open-loop/non-reactive/pseudo/closed-loop evaluation ladder, public benchmark interpretation, domain transfer caveats, and release artifacts. |
| Recurring ultrasonic proximity sensing loop | Ultrasonic Proximity Sensing Models promoted as a sensor foundation page covering pulse-echo time-of-flight, speed-of-sound compensation, beam/dead-zone geometry, echo confidence, cross-talk scheduling, safety-rated versus advisory use, domain transfer, and validation gates. |
| Recurring thermal radiometry loop | Thermal IR Radiometry First Principles promoted as a sensor foundation page covering emitted radiance, reflected background, atmospheric path, optics, microbolometer response, NUC tables, radiometric parameters, error budgets, domain transfer, and fusion-ready confidence. |
| Recurring fleet calibration operations refresh | Sensor Calibration Fleet Operations refreshed with calibration package lifecycle states, artifact manifest fields, versioned telemetry schema rules, drift remediation workflow, anti-patterns, rollback/quarantine discipline, and source links for Autoware, NVIDIA DriveWorks, ROS multisensor calibration, OCAMO, VCA SUMS/CSMS, Uptane, and OpenTelemetry. |
| Recurring infrastructure-aided localization loop | Infrastructure-Aided Localization promoted as a managed-site localization overview covering UWB anchors, fiducials, RFID/BLE, Wi-Fi RTT, magnetic maps, surveyed reflectors, 5G NR/mmWave positioning, measurement contracts, lifecycle controls, failure modes, and estimator handoff patterns. |
| Recurring cooperative source-hygiene refresh | CoInfra, HoloVIC, and V2X-ReaLO source routing refreshed: corrected CoInfra to arXiv 2507.02245, corrected V2X-ReaLO to arXiv 2503.10034, corrected HoloVIC CVF links, and refreshed V2XScenes/UrbanIng-V2X/V2XPnP/CoopScenes routing. |
| Recurring fiducial/corner localization loop | Fiducial and Corner Localization 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 patterns, and managed-site failure modes. |
| Recurring RKO-LIO loop | RKO-LIO 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. |
| 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 paradigms, tiling/throughput, post-processing, and static-transient point removal. |
| Recurring semantic-map dataset proxy loop | SemanticTHAB, WHU-Urban3D, CUS3D, and SUM/SUM Parts were added to the large-scale 3D segmentation benchmark landscape with urban-district, construction, mesh, and non-road transfer notes. |
| Recurring GridNet-HD utility-infrastructure loop | GridNet-HD was promoted as a LiDAR-image utility-infrastructure benchmark for pylon, cable, insulator, vegetation, soil/road, water, and building classes, then routed through the large-scale benchmark page, aggregated-map selector, taxonomy caution layer, perception audit, README, INDEX, methodology, and active source-watch query. |
| Recurring managed-site/facade proxy loop | Point Cloud City / Open3D-ML PCC and City-Facade were routed through the large-scale 3D segmentation benchmark page, aggregated-map hub, taxonomy design, perception audit, source registry, README, INDEX, and methodology as managed-building/public-safety and MLS facade/vertical-structure proxies. They support terminal-interior/frontage, label-unification, small safety-feature, and digital-twin taxonomy design, but are not direct airside acceptance datasets. |
| Recurring ZAHA facade-hierarchy loop | ZAHA was routed through the large-scale 3D segmentation benchmark page, aggregated-map hub, taxonomy design, perception audit, source registry, README, INDEX, and methodology as a WACV 2025 / TUM2TWIN MLS facade hierarchy benchmark with 601 M annotated points and LoFG2/LoFG3 classes for terminal-frontage and digital-twin taxonomy stress. |
| Recurring OpenLiDARMap georeferencing loop | OpenLiDARMap and the complementary FlexCloud branch were routed through the HD map-construction pipeline, aggregated-map source conditioning, source registry, README, INDEX, and methodology surfaces as map-prior / GNSS-assisted georeferenced point-cloud map-conditioning references, not semantic segmentation datasets or label sources. |
| Recurring LAMM multi-session map-merging loop | LAMM Multi-Session Point-Cloud Map Merging was promoted as the large-scale multi-session point-cloud map-merging branch after Uni-Mapper, then routed through dynamic map cleaning, map construction, aggregated-map segmentation, SLAM audit, source registry, README, INDEX, and methodology surfaces with explicit license/research-artifact caveats. |
| Recurring MapEval map-quality loop | MapEval Point-Cloud Map-Quality Evaluation was promoted as the direct point-cloud geometry QA gate after map merging/dynamic cleanup and before aggregated-map semantic segmentation, localization regression, or map publication. LEMON-Mapping remains watchlist/comparison until official code, license, or publication maturity improves. |
| Recurring MapEval QA-contract loop | The MapEval/source-geometry gate was tightened into a contract: metrics_evidence.qa_report_id now explicitly dereferences to a QA bundle carrying source_map_quality method, metric set, config hash, reference/no-reference evidence, alignment transform, threshold policy, failure-region digest, and gate status. The runtime-map example was also adjusted to Autoware's 20 m divided-point-cloud convention instead of a stale 100 m cell. |
| Recurring modality-aware architecture loop | Aggregated-Map Semantic Segmentation was refreshed with explicit colorized-cloud, train-time-distillation, image-dependent-fusion, and candidate-label lanes plus a backbone-by-modality deployment contract for sparse-conv, KPConv/RandLA, SPT, PTv3/Sonata, projection, and SSM/Mamba families. Companion pages now keep training-regime, tiling-formula, and single-scan/Orin concerns in their own scopes. |
| Recurring static-map lifecycle loop | Lifelong 3D map version control and detector-based potentially dynamic object removal were promoted into SLAM/map-lifecycle pages, then cross-linked into the aggregated-map segmentation, static-transient, dynamic-cleaning, policy, audit, README, and INDEX routes. |
| Recurring Uni-Mapper map-merging loop | Uni-Mapper was promoted as a dynamic-aware heterogeneous-LiDAR map-merging page, then cross-linked into dynamic cleaning, static-transient removal, aggregated-map segmentation, map construction, SLAM audit, open-source stack comparison, README, and INDEX routes. |
| Recurring Point-Cloud SSM backbone loop | Point-cloud Mamba/SSM backbones were promoted as one research-stage method-family page, then cross-linked into aggregated-map segmentation, LiDAR semantic segmentation, tiling/throughput, the perception audit, README, INDEX, and methodology routes. |
| Recurring semantic-map artifact contract loop | Aggregated-map segmentation, map construction, tile versioning, SLAM/map benchmark validation, map publication gates, and OTA compatibility were refreshed with a signed semantic-map manifest, taxonomy/model/config/calibration/evidence compatibility hash, Autoware runtime export contract, and loader smoke-test gates. |
| Recurring semantic-map schema contract loop | ../../schemas/semantic-map-manifest.schema.json, ../../schemas/runtime-map-contract.schema.json, and ../../examples/map-contracts/ now make the semantic-map manifest and runtime map contract machine-readable, example-backed, and checkable through npm run map-contracts:check. |
| Recurring semantic/SLAM prior contract loop | Semantic SLAM, neural/Gaussian SLAM surveys, dynamic 4D Gaussian SLAM, and the aggregated-map segmentation hub now route semantic-SLAM histograms, Gaussian motion priors, neural map features, and static/dynamic masks through manifest prior_inputs as QA or CRF evidence, not authoritative release labels. |
| Recurring semantic-map fleet telemetry loop | Map-hygiene monitoring, the perception-SLAM fleet data contract, and alert runbooks were refreshed with semantic-layer IDs, manifest/compatibility hashes, taxonomy IDs, schema URLs, unknown/coverage/safety-class/seam/churn telemetry, and semantic-map drift response. |
| Recurring open-vocabulary map-labeling bridge | Aggregated-map segmentation, taxonomy design, active labeling, replay mining, and model-governance pages were refreshed so ZOPP, SALT, OpenUrban3D, SAM4D, and similar outputs remain candidate labels until reviewer, taxonomy, QA, manifest, and release-evidence gates promote them. |
| Recurring semantic-map data-engine bridge | Data-engine, annotation-tool, active-labeling, synthetic-FOD, and flywheel pages were refreshed so VESPA, UniLiPs, LOSC, SALT, SAM4D, and OpenUrban3D outputs are stored as candidate or pseudo-labeled regions with prompt, checkpoint, calibration, consolidation, taxonomy, QA, manifest, and back-projection provenance before training or release use. |
| Recurring semantic-map source-hygiene loop | Aggregated-map segmentation, training paradigms, tiling/throughput, open-vocabulary LiDAR, LiDAR semantic segmentation, Mosaic3D, the perception audit, data-engine, and annotation pages were refreshed to time-scope UniLiPs and LOSC, remove unsupported SALT cost percentages, clarify OpenUrban3D, SAM4D, and VESPA source boundaries, and split SALT and SAM4D routing. |
| Recurring LOSC consolidation loop | LOSC was promoted as an atomic open-vocabulary LiDAR pseudo-label consolidation method and routed through aggregated-map segmentation, open-vocabulary detection, data-engine, annotation, audit, README, INDEX, and methodology surfaces. |
| Recurring map-hygiene protocol workflow loop | The canonical Airside Map Hygiene Ground Truth Protocol and its V&V companion were wired into aggregated-map QA, map construction, publication gates, operational monitoring, regulatory evidence, README, INDEX, and the perception audit. |
| Recurring aggregated-map selector loop | Aggregated-Map Semantic Segmentation now includes a compact proxy/input/training selector that routes airside, urban-district, non-road, calibrated LiDAR+image, existing-SLAM-map, corridor, and mesh/digital-twin cases to the right datasets, model family, and training regime. |
| Recurring FOD proxy-boundary loop | FOD/apron dataset pages, adverse-weather proxy routing, FOD validation, FOD-retention map cleaning, and LiDAR artifact-removal guidance now separate public screening/training proxies from target-airside acceptance evidence for airside dust, de-icing mist, steam, glycol film, wet-apron multipath, retroreflector bloom, and do-not-delete hazards. |
| Recurring local holdout-routing loop | Canonical map-hygiene labels, the V&V companion, false-deletion testing, FOD perception/retention validation, publication gates, and potentially dynamic object policy now route do-not-delete hazards and adverse-airside slices into locked local holdouts with raw/rejected evidence, reviewer disposition, and quarantine/ODD-restriction decisions. |
| Recurring BeautyMap dynamic-removal loop | BeautyMap was promoted as the binary-encoded ground-matrix branch for fast offline dynamic-point removal, then cross-linked into dynamic map cleaning, benchmarks, aggregated-map segmentation, static-transient removal, artifact removal, audits, README, INDEX, and methodology routes. |
| Recurring Raymoval dynamic-removal loop | Raymoval was promoted as the paper-backed azimuth-elevation raycasting branch for dynamic residual removal, with metric-lineage caveats and paper-only maturity captured before cross-linking through map cleaning, benchmarks, segmentation, static-transient removal, artifact removal, audits, README, INDEX, and methodology routes. |
| Recurring DynPurge/PRN triage loop | Five scouts verified DynPurge's IEEE RA-L 2025 / DBLP record and public pseudocode/result-image repository, then kept it as a watchlist row rather than an atomic page because runnable code, metric tables, and full evaluation details are not public. PRN was routed as a learned-restoration caveat for static-but-transient cleanup because current public evidence is paper-level and CARLA-oriented, with no source artifact mature enough for a safety-critical map-publication method page. |
| Recurring multi-agent neural/Gaussian SLAM refresh | Multi-Agent Neural and Gaussian SLAM was refreshed with CoMA-SLAM's AAAI 2026 distributed 2D Gaussian surfel branch, MAGS-SLAM's RGB-only branch, source-backed Coko-SLAM routing, and a corrected frontier/research maturity rating so the page remains useful for ML-SLAM research without implying LiDAR-primary production readiness. |
Active Next Queue
| Priority | Queue | Next atomic files to consider |
|---|---|---|
| P0 | Perception occupancy and radar | LinkOcc remains watchlist until open method details or code make an atomic page reviewable. GaussianFlowOcc, GaussTR, GS-Occ3D, EvOcc, M2-Occ, TEOcc, and GaussRender are already promoted or routed. |
| P0 | Sparse and end-to-end perception | DySS is routed through the existing sparse-query and streaming-temporal overviews until official code or broader adoption justifies an atomic page. Remaining sparse-query/e2e perception follow-ons stay eligible where source maturity justifies non-duplicative atomic pages. SparseDriveV2 and DiffusionDriveV2 are now refreshed through the existing planning/E2E benchmark pages rather than split into duplicate atomic pages; SparseBEV, SparseDrive, DiffusionDrive, DETR4D, ForeSight, SAM4D, Open3DTrack, and DriveBench are already covered or cross-linked. |
| P0 | Removal validation and adverse-weather datasets | Airside dust, de-icing mist, steam, glycol film, wet apron multipath, retroreflector bloom, and do-not-delete hazard labels remain direct-data gaps. Airport-FOD3S is now promoted as a synthetic FOD data-engine page, the FOD proxy-boundary note is covered across dataset/validation/removal pages, and local holdout routing is now covered across map-hygiene labels, false-deletion testing, FOD validation, publication gates, and potentially dynamic object policy. DSERT-RoLL, CMHT, SemanticSpray++, RADIATE, and LIDAROC are only public proxies, not direct airside obscurant datasets. Remaining next candidates are actual target-airport data collection, a new public/primary airside source, or a narrow capture-manifest template if operational fields become concrete; avoid another duplicate overview. DR-REMOVER is covered under localization/mapping dynamic-object removal, and ExelMap is covered under map-change maintenance; keep them cross-linked instead of creating duplicate perception pages. |
| P0 | Aggregated-map segmentation pipeline | Point-cloud Mamba/SSM backbones are now promoted and routed as a research-stage efficiency branch; the production artifact contract, semantic/SLAM prior-input contract, fleet telemetry handoff, offboard/open-vocabulary candidate-label promotion lane, data-engine/annotation bridge, map-hygiene ground-truth workflow, compact proxy/input/training selector, FOD proxy-boundary routing, and source-hygiene corrections for UniLiPs, LOSC, SALT, OpenUrban3D, SAM4D, and VESPA are now refreshed across semantic segmentation, taxonomy, active labeling, replay, governance, map construction, tile manifests, validation, publication, OTA compatibility, observability, data contracts, annotation tooling, synthetic-FOD expansion, and training flywheel pages. LOSC is now promoted as the atomic consolidation method. WHU-Urban3D, CUS3D, SemanticTHAB, SUM/SUM Parts, GridNet-HD, Point Cloud City / Open3D-ML PCC, City-Facade, ZAHA, OpenLiDARMap/FlexCloud source-map conditioning, LAMM multi-session map merging, Uni-Mapper heterogeneous-LiDAR map merging, MapEval point-cloud map-quality evaluation, MapEval QA-report dereference requirements, and the urban-district/non-road/utility-infrastructure/managed-site/facade proxy matrix are now covered. Next candidates: II-NVM as a normal-vector-assisted map-conditioning page, DQFormer as a query-based LiDAR panoptic segmentation page, or MS-Mapping as a multi-session mapping page only if a future pass can make the scope non-duplicative and license/source maturity is clear. LEMON-Mapping stays watchlist-only until official code, license, or publication maturity improves. |
| P0 | Static-but-wrong map lifecycle | Lifelong 3D map version control, detector-based potentially dynamic object removal, Uni-Mapper dynamic-aware heterogeneous-LiDAR map merging, LAMM multi-session point-cloud map merging, BeautyMap binary-matrix cleaning, Raymoval raycasting residual cleaning, the production artifact-contract refresh, semantic-map fleet telemetry, and open-vocabulary candidate-label promotion gates are now routed. DynPurge is now a watchlist-only timestamp-distribution cleaner for global-scale LiDAR maps: promote only if full text, runnable code, metric tables, or benchmark artifacts make it non-duplicative against ERASOR, Removert, MapCleaner, BeautyMap, Raymoval, FreeDOM, STATIC-LIO, and MOVES. PRN remains mention-only as learned point restoration for temporarily static objects because current public evidence is paper-level, CARLA-oriented, and not a publication-gate-safe replacement for map lifecycle, quarantine, or do-not-delete evidence. Do-not-delete/adverse-airside local holdout routing is now covered; revisit only when new target-airport data, a primary public airside source, or concrete capture-manifest fields appear. |
| P0 | SLAM robustness and benchmarks | CAO-RONet is now promoted as the source-mature learned 4D radar odometry page. Doppler Radar-LiDAR SLAM now folds Radarize, DRO, and Doppler-SLAM into one source-backed bridge page; create deeper Radarize, DRO, or Doppler-SLAM pages only if a non-duplicative implementation or benchmark gap appears. Radar RIO Correspondence and Uncertainty now routes radar_transformer, HKUST point-uncertainty RIO, continuous point-pose uncertainty, and UNRIO, so create deeper radar correspondence or point-uncertainty pages only if a narrower implementation gap appears. Radar Place Recognition now routes 4DRaL, SHeRLoc, TransLoc4D, 4D RadarPR, and TDFANet, so create deeper radar descriptor pages only if a non-duplicative implementation, reproducibility, or benchmark gap appears. Multi-Agent Neural and Gaussian SLAM now routes CoMA-SLAM, MAGS-SLAM, Coko-SLAM, and earlier collaborative neural/Gaussian branches as frontier research rather than a LiDAR-primary production stack; create separate CoMA/MAGS pages only if deeper reproduction or benchmark evaluation becomes necessary. GaussianFlow-SLAM remains watchlist-only: the RA-L 2026 paper and project page are available, but the official repository still contains only README/media and says code is under construction, so do not promote it until runnable code, configs, setup instructions, and evaluation scripts are public. Wheel/vehicle-motion factors and raw GNSS factor fusion are now promoted; create deeper wheel, visual-inertial-wheel, or learned wheel-kinematic pages only if the scope is not duplicative. Oxford Spires, IILABS 3D, SMapper-light, FusionPortableV2, and S3E are now matrix-covered; create dedicated pages only if a deeper dataset evaluation is needed. LaMAria, Hilti x Trimble 2026, ScaleMaster, SLAM Toolbox, NDT variants, HI-SLAM2, SEGS-SLAM, the neural/Gaussian survey page, Multi-Agent Neural and Gaussian SLAM, GV-iRIOM, GVINS/GLIO raw GNSS factor fusion, wheel/vehicle-motion factors, CAO-RONet, Doppler Radar-LiDAR SLAM, Radar RIO Correspondence and Uncertainty, radar place recognition, and radar-IMU online spatio-temporal calibration are now promoted or refreshed. |
| P0 | Sensor and calibration fundamentals | Online calibration drift response remains a watch item only when it is not duplicative of fleet calibration operations, active calibration experiment design, SOAC, RC-AutoCalib, or radar-inertial online calibration. Fiducial/corner localization is now promoted through the geometry foundation page; thermal IR radiometry and ultrasonic proximity models are promoted through sensor-foundation pages; calibration-bay fixtures are promoted, and active calibration experiment design covers the safe data-collection/observability bridge. |
| P0 | First-principles extensions | Robust-loss uncertainty follow-ons. Dataset-evaluation fundamentals are now promoted through the AV data-evaluation foundation page, constrained KKT/QP/SQP mechanics are now promoted through the constrained-solver foundation page, optical/scene-flow math is now promoted through the motion-field first-principles page, optimal experiment design for calibration is now promoted through the active calibration experiment-design page, covariance consistency under robust losses is now promoted through the robust-loss covariance page, and epipolar/homography foundations are now promoted through the two-view verification page. |
| P1 | Collaborative and infrastructure perception | GSCOOP remains watchlist-only as a collaborative Gaussian BEV semantic-segmentation follow-on after VOGS-CP; the ICCV 2025 paper/poster are available, but the stated official repository is still publicly inaccessible. CoDS remains watchlist-only pending code/project/final-venue artifacts; JigsawComm remains watchlist-only until its placeholder repo matures into a released implementation or venue-backed project. SpaCeFormer now covers source-mature indoor open-vocabulary 3D instance segmentation. QuantV2X, SparseCoop, TruckV2X, VOGS-CP, and EmbodiedScan/MMScan benchmark coverage are now promoted. |
| P1 | Alternative localization sensors | Benchmark/dataset follow-ons for degraded localization, source-mature radar odometry/place-recognition variants not already covered by the radar family pages, and deeper infrastructure-aid method pages only when they are not duplicative of the new infrastructure-aided localization overview. Wheel/LiDAR/IMU factor graphs now have a factor-family entry point through the wheel/vehicle-motion factors page plus CM-LIUW and semantic-LIW pages. |
Promotion Rules
- Prefer one file per method or technique when the user needs depth.
- Keep family synthesis in overview files; put method evidence in method files.
- Use primary sources first: papers, official project pages, official repos, standards, or vendor documentation.
- Every generic method page should state inputs, outputs, assumptions, failure modes, AV relevance, and Domain Fit across the relevant canonical domains. Domain-specific pages should state their primary ODD and include transfer notes when the idea generalizes.
- Every sensor/foundation page should connect measurement physics to perception, SLAM, mapping, validation, and operational monitoring.
- After every wave, update the relevant coverage audit, Research Index, README, and Methodology.
Source Audits
| Audit | Role |
|---|---|
| Active Frontier Source Registry | Lists source sites, native filters, query patterns, review cadence, canonical routing rules, and semi-automation boundaries for active frontier monitoring. |
| Perception Coverage Audit | Tracks perception methods, benchmarks, datasets, and robustness gaps. |
| SLAM Coverage Audit | Tracks SLAM, odometry, localization, backend, sensor-fusion, and benchmark gaps. |
| Knowledge Gap Backlog | Tracks cross-architecture gaps outside the dedicated perception and SLAM audits. |
| Cross-Architecture Knowledge Base Gap Fill | Turns the latest broad KB web sweep into source-backed promotion queues for safety, runtime, platform, foundations, autonomy evaluation, and deployment intelligence. |