Skip to content

Neural/Gaussian SLAM Surveys

Related docs: SLAM Method Library Overview, Photoreal City-Scale 4D Reconstruction, Volume Rendering, Radiance Fields, and Gaussian Splatting, Feed-Forward 3D Reconstruction and Splatting, SLAM3R and VGGT Foundation SLAM, Multi-Agent Neural and Gaussian SLAM, HI-SLAM2, and SEGS-SLAM.

Executive Summary

Neural and Gaussian SLAM has become too broad to manage as a single method family. The 2024-2026 literature now includes NeRF-style implicit maps, 3D Gaussian maps, feed-forward pointmap/foundation SLAM, multi-agent neural maps, semantic Gaussian maps, dynamic Gaussian SLAM, LiDAR-camera Gaussian mapping, and radar/Gaussian hybrids.

This page is a taxonomy and routing guide. Its job is to keep individual method pages organized and to make the production boundary explicit: neural/Gaussian maps are promising dense map representations, visual QA artifacts, simulation assets, and research localizers, but they are not yet a general replacement for validated multi-sensor localization.

The survey sources are useful for structure, not for deployment claims by themselves. Any performance or readiness claim should be checked against the individual method paper, code, dataset, and local validation.

Taxonomy

FamilyRepresentative local pagesMain representationPrimary valueMain deployment blocker
NeRF/implicit dense SLAMiMAP, NICE-SLAM, Co-SLAM / ESLAM, NeRF-SLAMImplicit neural fields, feature grids, SDF/radiance fieldsDense reconstruction and compact scene functionsRuntime, uncertainty, map editing, and metric integrity
First-wave 3DGS visual SLAMGS-SLAM and MonoGS, SplaTAM, Photo-SLAM3D Gaussian primitivesFaster differentiable rendering and photorealistic mapsStatic-scene and camera-fragility assumptions
Globally corrected RGB Gaussian SLAMSplat-SLAM, HI-SLAM2RGB-derived 3D Gaussian maps with global correctionLoop-consistent dense visual reconstructionMonocular scale, learned priors, and no physical sensor authority
Outdoor/foundation visual Gaussian SLAMS3PO-GS, SLAM3R and VGGT Foundation SLAMPointmaps, feed-forward geometry, GaussiansLarge-scale visual reconstruction and learned geometry priorsDomain shift, hallucination, and weak safety evidence
Structured visual Gaussian mappingSEGS-SLAM, Photo-SLAMStructured Gaussian anchors and appearance embeddingsBetter photorealistic mapping across camera modesRendering quality can diverge from pose integrity
Multi-sensor Gaussian SLAMGaussian-LIC, RMGS-SLAM, GS-LIVM, VIGS-SLAMGaussian maps constrained by LiDAR/camera/IMUMore metric outdoor mapping and visual QACalibration, sensor degradation, dependency complexity
Dynamic/radar Gaussian SLAMDynamic 4D Gaussian SLAM, RadarSplat-RIO, WildGS-SLAMTime-aware or radar-constrained GaussiansDynamic-scene and adverse-weather researchRobust dynamic-object lifecycle and certified uncertainty
Collaborative neural/Gaussian SLAMMulti-Agent Neural and Gaussian SLAM, COSMO-BenchShared neural/Gaussian maps or backend graph dataMulti-robot mapping and dense digital twinsCommunication, false inter-agent loops, and map consistency

Representation Choices

RepresentationGood atWeak atAV interpretation
NeRF / implicit fieldSmooth dense reconstruction and continuous renderingSlow optimization, hard map editing, hidden geometry failuresUseful for offline reconstruction and simulation, less practical for runtime pose
3D Gaussian splatsFast rendering and explicit primitivesUncertainty, dynamic-object ghosts, map lifecycle, primitive explosionStrong visual QA/simulation artifact if aligned to trusted trajectories
Pointmaps / feed-forward geometryFast dense visual priors and calibration-light reconstructionLearned-prior hallucination and scale ambiguityUseful for camera-log mining and dense map proposals, not authority
LiDAR/camera/IMU-constrained GaussiansMetric geometry plus photorealistic appearanceCalibration and multi-rate synchronization sensitivityMost AV-relevant neural/Gaussian direction, still research-stage
Dynamic/time-aware GaussiansReconstructing non-static scenesSeparating moving hazards from persistent map truthUseful for map-cleaning research, not yet runtime safety state

Production-Readiness Ladder

LevelDescriptionExamplesGate before promotion
Visual research mapperRGB/RGB-D Gaussian or NeRF map with trajectory outputGS-SLAM, MonoGS, SplaTAM, SEGS-SLAMReproduce paper metrics and document alignment policy.
Globally corrected visual mapperLoop or global correction updates dense map stateSplat-SLAM, HI-SLAM2Prove loop false-positive handling and scale behavior.
Foundation-prior mapperLearned pointmaps/depths initialize or regularize SLAMSLAM3R, VGGT-SLAM, S3PO-GSTest domain shift and hallucination against physical sensors.
Metric multi-sensor Gaussian mapperLiDAR/IMU/camera constrain Gaussian mapsGaussian-LIC, GS-LIVM, RMGS-SLAM, VIGS-SLAMValidate calibration, covariance, timing, and sensor-fault behavior.
Production localization authorityCertified map-frame pose with health and fallbackConventional scan-to-map localization, LIO/VIO/RIO plus HD mapRequires bounded latency, covariance consistency, map lifecycle, and safety evidence.

Most neural/Gaussian SLAM is currently in the first four levels. A system can be useful for production QA without being the production pose authority.

Evaluation Checklist

For every neural/Gaussian SLAM method, report:

  • trajectory ATE/RPE and alignment mode;
  • scale drift for monocular methods;
  • rendering metrics such as PSNR, SSIM, and LPIPS;
  • reconstruction accuracy/completeness against depth, LiDAR, or mesh truth;
  • map size, primitive count, memory, and runtime P95/P99;
  • loop-closure and relocalization false positives;
  • dynamic-object ghosting and static-map contamination;
  • calibration, timestamp, and sensor-degradation assumptions;
  • license and dependency constraints;
  • whether generated maps are used for visualization, simulation, QA, or live pose.

For autonomy, add physical-sensor disagreement and route-level safety gates. Good rendering is not enough.

Domain Fit

DomainUseful applicationsCaution
Indoor warehouse, terminal, hangar, labDense visual QA, digital twins, multi-robot inspection mapsDynamic workers/equipment and repetitive corridors need loop verification.
Road AV, outdoor campus, yard, port, constructionOffline reconstruction, map-change visualization, simulation assetsWeather, lighting, long range, and dynamic traffic require metric sensor checks.
Mining and agricultureTerrain/asset reconstruction under controlled capturesDust, vibration, vegetation, and low texture stress visual priors.
Delivery robotsSidewalk visual reconstruction and route reviewSmall platforms have compute, vibration, and privacy constraints.
AirsideHangar/terminal reconstruction, stand visual QA, simulation layersOpen tarmac, reflective aircraft, wet pavement, night floodlights, GNSS multipath, and moving GSE demand physical-sensor authority.

Routing Rules

  1. Put one-method evidence in an atomic method page.
  2. Use this page for survey/taxonomy updates and family routing only.
  3. Do not copy survey tables or broad performance summaries without checking the original method paper.
  4. Treat project pages and repositories as artifact evidence, not deployment evidence.
  5. If the method relies on RGB-only input, separate rendering quality from localization integrity.
  6. If the method uses LiDAR/camera/IMU, inspect calibration, timing, and failure monitoring before raising deployment relevance.
  7. If a method is dynamic or collaborative, require outlier, communication, and map-consistency tests before treating it as more than research.
  8. If a neural, Gaussian, or semantic-SLAM map is exported to Aggregated-Map Semantic Segmentation, declare it as manifest prior_inputs: pose provenance, calibration hash, dynamic-mask policy, rasterization policy, uncertainty summary, and downstream use are required before the prior can affect CRF unaries, QA disagreement maps, or human-review routing.

Open Follow-Ups

  • Radar and online calibration: GV-iRIOM and radar-IMU spatio-temporal calibration deserve a separate physical-sensor robustness pass.
  • Occupancy and Gaussian perception: GS-Occ3D and GaussTR belong in the perception method library, not this SLAM page.
  • Foundation visual SLAM splits: VGGT-SLAM++, ViSTA-SLAM, GaussianFlow-SLAM, and MegaSaM should remain in or split from the foundation-SLAM page only when source maturity and reader demand justify it.
  • Collaborative dense maps: use COSMO-Bench for backend benchmarking, then compare dense map fusion through Multi-Agent Neural and Gaussian SLAM.

Practical Recommendation

Use neural/Gaussian SLAM as a map-representation and reconstruction research layer. For production autonomy, keep the authoritative pose in a conservative, testable localization stack and use neural/Gaussian maps for QA, simulation, inspection, and research until uncertainty, lifecycle, dynamics, and failure monitoring mature.

Sources

Public research notes collected from public sources.