Skip to content

SLAM Decision Matrix for AV, Indoor, and Outdoor Systems

This file is a practical selection matrix for choosing SLAM, odometry, and localization methods by operating domain. It is intentionally biased toward deployability: sensor availability, timing, map product, failure detection, license, and maintenance matter as much as benchmark error.

Related areaLinkUse in this decision matrix
LiDAR front-end detailsLiDAR SLAM AlgorithmsMethod-level evidence for KISS-ICP, LIO-SAM, FAST-LIO2, Faster-LIO-style voxel LIO, CT-ICP, and Point-LIO.
Runtime map localizationProduction LiDAR Map LocalizationUse when a validated map exists and online SLAM should not be the global truth source.
Loop closure and recoveryLiDAR Place Recognition and Re-LocalizationRequired when the matrix recommends loop closure, multi-session mapping, or kidnapped-robot recovery.
Survey-to-map workflowMap Construction PipelineUse when choosing a stack for airport onboarding, map merging, or QA.
Multi-sensor fusionRobust State Estimation Multi-SensorUse when the decision depends on RTK, IMU, wheel, covariance, and fallback behavior.
Factor graph backendGTSAM Factor GraphsUse when a method needs IMU preintegration, scan-matching factors, loop closure, GCP factors, or smoothing.
Dense/neural mappingGaussian Splatting for DrivingUse for future dense maps, inspection, simulation, and semantic QA rather than primary safety pose.

Fast Selection Table

Operating contextPrimary recommendationSecondary baselineRecovery/loop closureDo not choose as primaryReasoning
Airport airside AV, mapped ODDScan-to-map VGICP/NDT + GTSAM fusionKISS-ICP odometry for independent fallbackScan Context/MinkLoc3D + ICP/NDT verificationPure online SLAM as runtime global truthThe environment is mapped, safety-critical, and georeferenced; bounded drift beats live map growth.
Airport survey mappingFAST-LIO2 or GLIM + loop closure + GCP factorsKISS-ICP or CT-ICP validation runLIO-SAM style graph or KISS-SLAMVisual-only SLAMSurvey needs geometric accuracy, map consistency, and independent checks against IMU/extrinsic mistakes.
Urban road AV, mapped ODDHD-map localization with LiDAR/radar/GNSS/INS fusionAutoware NDT or MOLA localizationPlace recognition and map-change detectionMonocular SLAM-onlyRoad AV localization is a map-matching and state-estimation problem, not just local SLAM.
Warehouse AGV, flat floorSLAM Toolbox or Cartographer 2D3D LiDAR odometry if racks/ramps matterAMCL-style global localization, reflectors/AprilTagsHeavy 3D LIO unless needed2D maps are sufficient, cheap, explainable, and easy to integrate with Nav2.
Multi-floor indoor or construction3D LiDAR-inertial SLAMRTAB-Map if RGB-D/cameras are strongMulti-session loop closure2D grid-only SLAMStairs, ramps, shafts, and partial floors break planar assumptions.
Outdoor campus/service robotKISS-SLAM, LIO-SAM, or GLIMKISS-ICP local odometryLong-term place recognitionOne-session map with no maintenance planCampus changes seasonally and structurally; long-term relocalization matters.
UAV or fast handheld scannerPoint-LIO, FAST-LIO2, or OpenVINS depending payloadVINS-Fusion visual-inertial baselineVisual or LiDAR place recognitionSlow scan-to-scan ICP with no motion modelAggressive motion and vibration require IMU-aware deskew and high-rate state output.
RGB-D indoor reconstructionRTAB-Map, NICE-SLAM, SplaTAMORB-SLAM3 RGB-D modeDBoW/visual loop closureLong-range outdoor LiDAR stacksDense geometry and appearance matter more than long-range AV robustness.

Score Legend

ScoreMeaningDeployment interpretation
5Strong defaultStart here unless a hard constraint blocks it.
4GoodViable with standard engineering and validation.
3ConditionalWorks when assumptions match; needs careful testing.
2Research or fallbackUseful as baseline, diagnostic, or constrained deployment.
1Poor fitUsually wrong for this domain.

Domain-Method Matrix

Method familyAirside AV runtimeAirside survey mappingRoad AVIndoor warehouseMulti-floor indoorOutdoor campusUAV/handheldMain reason
Prebuilt-map LiDAR localization5354342Best when a validated map exists; not a map builder.
LiDAR-only odometry3433343Simple and independent, but drift is unbounded.
LiDAR-inertial odometry4543555Best real-time geometric front-end when timing is good.
Factor-graph LiDAR SLAM3533554Adds loop closures, GPS/GCP factors, and multi-session consistency.
2D LiDAR graph SLAM1115221Excellent planar indoor fit; wrong abstraction for 3D AV geometry.
Visual SLAM2223334Useful with rich texture and cameras; fragile in glare, darkness, weather.
Visual-inertial odometry2233435Strong for drones/handheld; less robust than LiDAR in airside geometry.
RGB-D dense SLAM1114312Range-limited indoor mapping, not outdoor AV localization.
Radar odometry/SLAM3241131Weather-robust outdoor auxiliary; not yet a complete general replacement.
Gaussian/neural SLAM1213322Strong for dense maps and QA; immature for certified pose and uncertainty.

Airside AV Detailed Matrix

RequirementWeightScan-to-map VGICP/NDTFAST-LIO2LIO-SAMKISS-ICPORB-SLAM3Gaussian SLAM
Bounded global drift in mapped airport5523211
Robustness to night/glare/weather5555522
Open apron degeneracy handling5434311
Multi-LiDAR compatibility4533512
Certifiable diagnostics4434431
Map update discipline4523222
RTK/GCP integration4535221
Real-time Orin feasibility4454532
License/deployment risk3425522
Recommendation-Runtime primarySurvey/front-endSurvey/global graphFallback/validationAuxiliary onlyQA/research

Indoor Decision Matrix

Indoor conditionRecommended stackWhyAdd-onsWatch-outs
Flat warehouse, differential or Ackermann robotSLAM Toolbox + Nav2 + wheel odometryOccupancy grid and 2D pose graph are simple and adequate.AMCL, reflectors/AprilTags, map zonesRacks and pallets create aliasing; update maps deliberately.
Warehouse with tall racks, mezzanine, ramps3D LiDAR-inertial SLAM or Cartographer 3D2D maps lose vertical structure and can confuse floors.Floor segmentation, elevator/stair constraintsRepeated aisles require strong relocalization verification.
Construction site or undergroundGLIM, FAST-LIO2, Point-LIOLow light, dust, and non-planarity favor LiDAR plus IMU.Multi-session mapping, robust kernels, dynamic filteringSensor protection and sync are often bigger risks than algorithm choice.
Office/RGB-D mappingRTAB-Map or RGB-D visual SLAMDense colored maps and object context are useful.Loop closure, TSDF/surfel exportSunlight and glass degrade depth cameras.
AR/headset room-scale trackingOpenVINS, ORB-SLAM3, visual-inertial stackCameras and IMU are the native sensors.Relocalization maps, anchorsMonocular scale and initialization need strong handling.

Outdoor Decision Matrix

Outdoor conditionRecommended stackWhyAdd-onsWatch-outs
Mapped road or airport ODDScan-to-map localization + state estimatorExisting map should bound drift.GNSS/INS, wheel odometry, map-change detectionMap staleness and dynamic-object residuals.
Unmapped survey driveFAST-LIO2 or GLIM + loop closuresBuilds accurate geometry quickly.KISS-ICP validation, GCP/RTK factorsDirect LIO map may still drift without loops/anchors.
Long-term campus routeKISS-SLAM, LIO-SAM, GLIMNeeds revisits, loops, and changing-season robustness.NCLT/Oxford RobotCar style validationVegetation, snow, construction, and traffic change the map.
Tunnel/urban canyonLiDAR-inertial plus wheel/vehicle kinematicsGNSS weak; geometry and IMU dominate.Radar, map priors, loop closuresTunnels can be longitudinally degenerate and repetitive.
Rain/snow/fog test vehicleLiDAR/radar/GNSS fusionRadar and GNSS reduce dependence on degraded LiDAR/camera.Boreas/Oxford long-term benchmarkNo single sensor is enough across all weather.

Sensor Availability Matrix

Sensors availableGood method familiesCandidate pagesMinimum extra checks
2D LiDAR + wheel2D graph SLAM, AMCL/localizationSLAM Toolbox, CartographerWheel scale, laser extrinsic, scan rate, planar assumption
3D LiDAR onlyLiDAR odometry/SLAMKISS-ICP, KISS-SLAM, CT-ICPDegeneracy, loop closures, deskew model
3D LiDAR + IMULIOFAST-LIO2, Faster-LIO family, Point-LIO, LIO-SAMTime sync, IMU noise model, extrinsics, saturation
Multi-LiDAR + IMU + wheelsProduction AV localization/fusionProduction LiDAR Map Localization, Robust State Estimation Multi-SensorPer-sensor extrinsics, covariance, hot-path memory, fault isolation
Stereo/RGB-D + IMUVIO/RGB-D SLAMORB-SLAM3, OpenVINS, VINS-Fusion, RTAB-MapExposure/blur, camera-IMU calibration, feature count
Radar + LiDAR/IMUAll-weather localization researchRadar odometry, radar place recognition, LiDAR-radar fusionRadar calibration, multipath, Doppler models

Compute and Integration Matrix

ConstraintPreferAvoidNotes
NVIDIA Orin with CUDA availableGPU VGICP, GLIM, gtsam_points factorsCPU-only algorithms that cannot meet multi-LiDAR throughputCUDA improves scan matching, but certification still needs bounded timing and fallbacks.
Raspberry Pi / low-power CPUKISS-ICP, 2D SLAM, light VIOHeavy neural/Gaussian SLAMKeep map bounded and downsample aggressively.
ROS 2 production stackAutoware localization, SLAM Toolbox, MOLA, GLIM ROS 2ROS 1-only research stacks unless wrappedROS version is often a schedule driver.
GPL avoidance for closed deploymentMIT/BSD/Apache stacks, in-house scan matcherGPL-2/GPL-3 libraries as linked product codeVerify legal interpretation before product integration.
Need factor-level fusionGTSAM-based pipelinesBlack-box pose output onlyFactor graphs expose residuals, covariance, robust kernels, and graph diagnostics.
Need explainable safety caseClassical scan matching + explicit diagnosticsNeural-only pose estimationExplainable residuals are easier to gate and audit.

Decision Tree

QuestionIf yesIf no
Is there a validated, current map of the operating area?Use scan-to-map localization as the runtime primary and SLAM only for map maintenance/fallback.Use SLAM/odometry to build an initial map, then convert to localization mode.
Is the vehicle safety-critical in a bounded ODD?Require covariance, fault detection, relocalization, and map version control.A research SLAM stack may be acceptable for prototyping.
Is the environment mostly planar and indoor?Evaluate 2D SLAM first.Use 3D LiDAR/LIO or visual-inertial depending sensors.
Does the system experience aggressive motion, vibration, or spinning LiDAR distortion?Use IMU-aware or continuous-time methods.A simple LiDAR-only ICP baseline may be enough.
Are repeated structures a core hazard?Add place-recognition verification, zone priors, and robust loop closure gates.Simpler loop closure may be acceptable.
Is the map product part of an HD-map pipeline?Select methods that export trajectories, submaps, graph constraints, and QA metrics.A black-box pose stream can still be useful for navigation demos.
GoalShortlistWhy this set
Airside map survey proof-of-conceptFAST-LIO2, KISS-ICP, LIO-SAM, GTSAM GCP factorsTests LIO accuracy, LiDAR-only independence, and graph correction.
Airside production localization prototypeGPU VGICP/NDT, Autoware NDT diagnostics, GTSAM/iSAM2, Scan Context recoveryMatches the production split: map localization, state estimation, recovery.
Indoor warehouse productSLAM Toolbox, AMCL/Nav2, reflectors/AprilTags, optional RTAB-MapFastest route to a reliable 2D navigation product.
Construction/underground mappingGLIM, FAST-LIO2, Point-LIO, Hilti benchmarkHandles 3D geometry, dark spaces, platform diversity, multi-session mapping.
Camera-first robotics researchORB-SLAM3, OpenVINS, VINS-Fusion, EuRoC/TUM-VIStrong visual/VIO baselines and standard datasets.
Dense map/semantic QA researchRTAB-Map, SplaTAM, NICE-SLAM, Gaussian Splatting for DrivingFocuses on reconstruction quality and appearance, not just pose.

Sources

Public research notes collected from public sources.