Artificial Intelligence-Enhanced Visual–Inertial Navigation for Autonomous Lunar Surface Mobility: Systems, Datasets, Methods, and Open Challenges

Abstract

Reliable onboard state estimation plays an important role in autonomous lunar surface mobility, given the lack of conventional global navigation satellite system (GNSS) coverage and the challenges of variable illumination and terrain texture. This review synthesizes model-based, learning-based, and hybrid visual–inertial odometry (VIO) approaches and assesses their relevance to lunar surface operations. The reviewed studies suggest that learned components can improve feature extraction, matching, depth estimation, and measurement confidence in selected datasets and operating conditions. However, representative lunar validation and uncertainty calibration remain limited. Given these limitations, hybrid architectures integrating learned components with explicit model-based estimators offer a comparatively testable design direction. Reduced lunar gravity alone does not render monocular scale unobservable; practical scale–bias performance also depends jointly on motion excitation, inertial quality, visual geometry, calibration, and estimation-window length. Stereo, wheel, crater, and map cues can provide additional metric information but introduce source-specific risks involving calibration, slip, association, and registration. Within the resources surveyed, we found no publicly available flight-derived lunar traverse dataset combining synchronized camera–IMU measurements, independent trajectory ground truth, and onboard timing. Evidence-based priorities include representative lunar validation, uncertainty-aware integration of learned components, consistency monitoring across metric sources, explicit estimator health and fallback mechanisms, and target-hardware resource evaluation.

1. Introduction

Planned and ongoing lunar programs increasingly call for longer and more autonomous surface operations than the Apollo traverses [1,2,3,4,5,6]. Communication delay, intermittent Earth visibility, limited relay coverage, and operational constraints imposed by extreme temperatures and variable illumination increase the importance of reliable onboard navigation. These challenges may be particularly significant within deep crater basins and on the lunar far side, where terrain occlusion can interrupt direct communication and continuous orbital-relay coverage cannot always be assumed.

Lunar surface vehicles operate without a terrestrial global navigation satellite system (GNSS), and may have incomplete orbital priors [7], necessitating reliable onboard state estimation for mobility. Visual–inertial fusion is an attractive state-estimation baseline because cameras provide geometric measurements while an inertial measurement unit (IMU) supplies high-rate motion information when visual constraints weaken. Established terrestrial approaches include multi-state constraint Kalman filter (MSCKF) filtering [8,9] and sliding-window or factor-graph estimation [10,11,12]. Here, visual–inertial navigation refers to onboard state estimation from visual and inertial measurements: visual–inertial odometry (VIO) provides local motion estimates, as shown in Figure 1, whereas visual–inertial simultaneous localization and mapping (VI-SLAM) additionally maintains a map and may include loop closure or global constraints. Map-aided localization and auxiliary metric cues are included only when they constrain this estimator. Mission-level planning, trajectory generation, and rover control are excluded. Terrestrial performance does not establish performance over the full lunar operating envelope. Illumination, texture, motion, sensor configuration, and size, weight, and power (SWaP) limits can affect different parts of the pipeline in different regimes. This review therefore examines artificial intelligence (AI)-enhanced VIO for lunar surface mobility across the processing pipeline, from visual and inertial measurement formation to tightly coupled estimation, metric-scale integrity (defined in Section 4.1), robustness, validation evidence, and onboard implementation constraints. It synthesizes the methods and evidence reported in the literature and identifies the remaining research gaps.

Figure 1. Representative imagery for lunar surface mobility and onboard navigation. (a) Lunar surface in a representative operational setting; (b) rover-navigation view illustrating onboard perception, localization, or traverse execution.

Three groups of lunar operating stressors organize the review. First, non-Lambertian regolith reflectance, albedo variation, and hard sunlit–shadow transitions can violate standard photometric models [13,14]. Depending on terrain, viewpoint, exposure settings, and the algorithm, these effects may reduce repeatable feature measurements or shift learned components away from their training distributions or datasets; they should not be interpreted as uniform defects across all lunar imagery. Second, some lunar motion and sensing conditions may weaken metric-state estimation. Lunar gravity is approximately $g_{Moon}=1.625\hspace{0.17em}{\mathrm{m}/\mathrm{s}}^2$, but reduced gravity alone does not make metric scale unobservable. Instead, practical scale–bias estimation depends jointly on rover motion excitation, inertial-sensor quality, visual geometry, calibration, and the length of the estimator’s sliding window. Calibrated stereo cameras, wheel odometry, crater catalogs, and orbital maps can provide additional metric constraints, but each introduces its own failure mechanisms. Weak disparity, calibration drift, wheel slip, incorrect landmark association, and map-registration error can make an otherwise metric measurement intermittent, biased, or inconsistent with the local VIO estimate. Permanently shadowed regions (PSRs) create an additional configuration-dependent challenge because passive visible-light measurements may become weak or unavailable [15]. Third, candidate mitigation methods impose different size, weight, and power (SWaP) requirements. Public mission descriptions do not support a single SWaP allocation that applies to all lunar vehicles. Deployment claims should therefore be tied to a specified sensor configuration, onboard processor, image resolution, numerical precision, update rate, and duty cycle [16,17,18].

Table 1 maps lunar operating stressors onto the VIO processing pipeline. Lunar photometry primarily affects the availability and reliability of visual measurements. These measurements, together with inertial and auxiliary metric cues, determine whether the estimator can maintain metric-scale integrity. Residuals, finite-horizon conditioning, sensor health, and source disagreement then support robust gating, reweighting, fallback, and failure reporting inside the navigation estimator. Within this pipeline, AI is used as an umbrella term for learning-based components and includes both modular and end-to-end approaches; component-level discussions throughout use the specific architecture or algorithm (e.g., learned features, neural depth prediction, or semantic segmentation networks). Current lunar evidence more strongly supports integrating testable learned components with explicit model-based estimation.

Table 1. Key lunar operating stressors, estimator impacts, required responses, and representative approaches reviewed. Listed approaches are not necessarily lunar-validated; evidence boundaries for representative systems are given in Section 6.1 and Section 6.3.

No.	Lunar Stressor	Estimator Impact	Required Response	Section	Representative Approaches Reviewed
1	Reflectance and albedo variation	Photometric instability; domain shift	Robust features; geometric/semantic anchors	Section 3	Learned local features and matchers [19,20]; semantic crater and landmark detectors [21,22,23,24,25,26]
2	High dynamic range (HDR); hard shadows	Saturation; texture loss; shadow-edge outliers	Exposure control; visual-health checks	Section 3 and Section 5	Low-light and PSR perception studies [27,28]; event-camera datasets [29,30]; terrestrial event- and thermal-based VIO systems [31,32,33,34,35,36,37]; no lunar-validated HDR front end identified
3	Slow motion; low gravity; finite-grade inertial measurement unit (IMU)	Weak monocular scale–bias conditioning	Excitation and uncertainty checks	Section 4	Observability and conditioning analysis [38,39]; observability-consistent estimators [40,41,42,43]; initialization and online calibration [44,45,46]; degeneracy-aware updates [38,47,48]; no lunar-validated method identified
4	Permanently shadowed regions (PSRs); intermittent vision	Missing visual updates; drift growth	Multimodal fallback; safe handover	Section 4 and Section 5	ShadowNav [49]; map- and DEM-aided localization [49,50,51,52]
5	Size, weight, and power (SWaP) limits	Restricted model, frame rate, and window size	Resource-aware scheduling and validation	Section 3, Section 4 and Section 5	Lightweight learned features [20]; sparse-update VIO [53]
6	Low-texture or repetitive terrain	Fewer or poorly distributed reliable correspondences	Coverage-aware matching; match-quality checks	Section 3	LuVo [54]; Lo-SLAM [55]; DEM-anchored SLAM [52]

Available evidence does not reproduce the complete operating envelope in a single benchmark. Terrestrial datasets [56,57,58], planetary analogs and facilities [59,60,61,62,63,64,65,66,67,68,69,70], and simulators each represent only part of the problem. For example, SPICE-HL3 [63] provides synchronized stereo-inertial, wheel-odometry, and illumination information, along with precise ground-truth sequences under controlled high-latitude lunar-analog conditions; however, it does not constitute lunar flight telemetry. This review identified no publicly available, flight-derived lunar surface dataset that combines rover camera–IMU measurements, independent trajectory ground truth, and onboard timing information. Accordingly, the modality and limitations of the supporting evidence are specified throughout the front-end, estimator, robustness, and benchmarking sections. This review focuses on post-landing visual–inertial state estimation for surface mobility. Powered descent, orbital operations, pure remote sensing, mission planning, and motion control lie outside its scope. Monocular, stereo, map-aided, wheel-assisted, learned, and model-based methods are considered only if they contribute to the formation of visual–inertial measurements or to state estimation. The review is organized around three interconnected technical layers. Front end: Which model-based, learning-based, and hybrid front-end designs improve the availability and reliability of visual measurements under lunar illumination, texture, geometry, and SWaP constraints? Estimator: How should visual, inertial, stereo, wheel, crater, and map information be fused to maintain observable, well-conditioned, accurate, and temporally continuous metric state estimates? Robustness and failure handling: Which residual tests, conditioning measures, robust costs, uncertainty models, sensor-health checks, and fallback mechanisms have credible foundations for lunar VIO? The discussion of robustness is confined to mechanisms within the VIO estimator or at its measurement interfaces. Mission planning, trajectory generation, rover control, and higher-level autonomous decision-making remain outside the scope of this review.

This review provides a VIO-centered system synthesis rather than proposing a new navigation algorithm. It traces the evolution of post-landing lunar surface localization from controlled external aids and dead reckoning to visual odometry, map- and crater-aided localization, and modern tightly coupled VIO and VI-SLAM. The technical discussion is organized along the processing chain, progressing from visual and inertial measurement formation to estimator architecture and metric-scale integrity, and then to robustness and failure handling. Representative model-based, learning-based, and hybrid systems are linked to their corresponding evidence modalities, while datasets and facilities are compared according to the claims they can substantiate. The resulting synthesis emphasizes benchmark validity, robust VIO design, metric-source compatibility, and flight-relevant implementation. Figure 2 presents the overall structure of the review and the relationships among its sections.

Figure 2. Overall structure and relationships between different sections. Section numbers in parentheses indicate where each topic is reviewed.

2. Background, AI-Enhanced VIO Framework, and Review Scope

2.1. Evolution of Lunar Surface Navigation

Lunar surface localization has evolved through overlapping and complementary architectures rather than through a strictly sequential replacement of previous approaches. In controlled laboratory and proving-ground systems, artificial fiducial or barcode markers, beacons, or surveyed targets are often employed to provide repeatable absolute position information. These aids are valuable for calibration and algorithm verification, but dependence on pre-installed infrastructure limits their relevance to unconstrained lunar traverses. The limited scalability of infrastructure-dependent approaches has motivated the development of onboard dead reckoning and visual odometry for planetary vehicles. Wheel encoders and inertial propagation provide high-rate local motion but accumulate errors due to bias, slip, and sinkage. Stereo visual odometry adds geometric motion constraints and has flight heritage on Mars [71], while lunar surface localization has also used stereo imagery, terrain reconstruction, and map-relative constraints [16]. Field trials and mission analyses provide additional evidence for rover visual motion estimation and localization under analog and lunar operating conditions [68,72,73,74,75]. Integrated analog systems, including Seeker and related planetary-rover field trials, further demonstrate long-range autonomous navigation at the system level [68,76].

Landmark- and map-aided methods address long-term drift by associating rover observations with known terrain, crater catalogs, or orbital products [49,50,51,52,77]. Their absolute information is intermittent and depends on viewpoint compatibility, catalog completeness, and registration quality. Modern VIO and VI-SLAM systems can incorporate these cues as optional factors while tightly coupling camera with IMU measurements for local state estimation. Local-to-orbital registration using boulders, imagery, light detection and ranging (LiDAR), or elevation maps demonstrates several mechanisms for periodically bounding local odometric drift [78,79,80,81,82,83]. More recent studies have investigated learned local features, semantic landmark extraction, monocular depth priors, measurement confidence, and health classification within lunar VIO pipelines [19,20,52,53,54,84,85,86,87,88,89,90,91,92,93,94,95].

Figure 3 summarizes this evolution. The sequence should not be interpreted as a claim that VIO makes wheel, map, crater, or external-reference methods obsolete. A practical lunar navigator may combine complementary information sources according to terrain, illumination, vehicle configuration, and resource availability.

Figure 3. Evolution of post-landing lunar surface localization toward AI-enhanced visual–inertial navigation. The key stages progress from controlled aids and dead reckoning to visual odometry and map- or landmark-aided localization, and finally to hybrid lunar VIO using learned features, semantics, confidence, or depth priors. The sequence is architectural rather than strictly chronological: wheel, crater, map, calibration, and estimator-health cues remain complementary.

2.2. Mathematical Formulation of Lunar VIO

A visual–inertial navigator estimates platform motion by combining the complementary bandwidth and failure characteristics of cameras and an IMU. A representative keyframe state is

$$x_k=\left[{\mathbf{R}}_{b_k}^W,{\mathbf{P}}_{b_k}^W,{\mathbf{V}}_{b_k}^W,{\mathbf{b}}_a,{\mathbf{b}}_{\omega },{\mathit{\theta }}_{\mathrm{c}\mathrm{a}\mathrm{l}}\right]$$

(1)

where $W$ denotes the world frame and $b$ the body (IMU) frame; ${\mathbf{R}}_{b_k}^W$, ${\mathbf{P}}_{b_k}^W$, and ${\mathbf{V}}_{b_k}^W$ are body attitude, position, and velocity in the world frame; ${\mathbf{b}}_a$ and ${\mathbf{b}}_{\omega }$ are the accelerometer and gyroscope biases, respectively; and ${\mathit{\theta }}_{\mathrm{c}\mathrm{a}\mathrm{l}}$ denotes estimated camera–IMU calibration or timing parameters. The continuous-time IMU measurement model is conventionally written as

$${\widehat{\mathsf{\omega }}}_t={\mathsf{\omega }}_t+{\mathbf{b}}_{{\mathsf{\omega }}_t}+{\mathbf{n}}^{\omega }$$

(2)

$${\widehat{\mathbf{a}}}_t={\mathbf{a}}_t+{\mathbf{R}}_W^t\mathbf{g}\mathrm{ᵂ}+{\mathbf{b}}_{{\mathbf{a}}_t}+{\mathbf{n}}^a$$

(3)

where ${\widehat{\mathsf{\omega }}}_t$ and ${\widehat{\mathbf{a}}}_t$ denote the measured angular velocity and acceleration, respectively; ${\mathsf{\omega }}_t$ and ${\mathbf{a}}_t$ are the true angular velocity and linear acceleration in the body frame; ${\mathbf{R}}_W^t$ is the rotation matrix that transforms the world frame to the body frame—here and throughout, subscripts and superscripts on R and T denote the source and target frames, respectively, and a superscript $t$ denotes the IMU body frame at time $t$; gᵂ = [0, 0, g]^T denotes the gravity term in a z-up world frame, with the sign convention under which (3) gives the specific force sensed by the accelerometer (i.e., the negative of the gravitational acceleration vector), while ${\mathbf{n}}^{\omega }$ and ${\mathbf{n}}^a$ denote the corresponding measurement noise. For a landmark $P_j$ observed by camera $i$, a calibrated visual measurement can be represented as

$${\mathbf{z}}_{i,j}=\pi \left({\mathbf{T}}_b^C{\mathbf{T}}_W^b{P}_j\right)+n_{i,j}^z$$

(4)

where $\pi \left(·\right)$ is the camera projection model; $T_W^b$ and $T_b^C$ denote the world-to-body and body-to-camera (extrinsic) transforms, respectively; and $n_{i,j}^z$ is the visual measurement noise. Feature-based systems form reprojection residuals from (4), whereas direct systems form photometric residuals. A fixed-lag smoother or factor graph estimates the state by minimizing a robust objective of the form:

$$X^{*}=\underset{X}{argmin}\left[{‖r_{\mathrm{p}\mathrm{r}\mathrm{i}\mathrm{o}\mathrm{r}}‖}^2+{\sum }_{\mathit{k}}\left({‖r_{\mathrm{I}\mathrm{M}\mathrm{U},k}‖}^2\right)+{\sum }_{i,j}\rho \left({‖r_{\mathrm{v}\mathrm{i}\mathrm{s},ij}‖}^2\right)+{\sum }_q\left({‖r_{\mathrm{a}\mathrm{u}\mathrm{x},q}‖}^2\right)\right]$$

(5)

where $X$ collects the keyframe states $x_k$ over the estimation window; $r_{\mathrm{p}\mathrm{r}\mathrm{i}\mathrm{o}\mathrm{r}}$ and $r_{\mathrm{I}\mathrm{M}\mathrm{U},k}$ are the prior and inertial residuals, respectively; $r_{\mathrm{v}\mathrm{i}\mathrm{s},ij}$ is the visual reprojection residual formed from (4); $\rho \left(·\right)$ limits the influence of visual outliers; and $r_{\mathrm{a}\mathrm{u}\mathrm{x},q}$ may encode stereo, wheel, crater, digital elevation model (DEM), or other configuration-dependent constraints. Filtering methods represent the same information recursively through state propagation and measurement updates. Equations (1)–(5) expose the central review logic: Learned components may improve measurement quality, but metric integrity and uncertainty still depend on the estimator, calibration, motion excitation, and compatibility of all metric sources. In AI-enhanced lunar VIO, a learned module may propose a measurement, prior, covariance, or gating score, but its output should be incorporated into an explicit estimation structure with calibrated uncertainty, geometric or statistical validity checks, and a defined response to unreliable inputs.

2.3. Model-Based, Learning-Based, and Hybrid VIO Architectures

This review classifies VIO approaches into three architectural categories according to the role of learned models in measurement formulation and state estimation: model-based, learning-based, and hybrid architectures. These categories describe differences in system design rather than technological maturity.

Model-based VIO relies primarily on explicit camera and inertial measurement models, handcrafted or direct visual residuals, and filtering or optimization with physically interpretable states and uncertainty [8,9,10,11,12]. Its advantages include physically interpretable state and residuals, explicit representations of measurement uncertainty, and established methods for calibration and observability analysis. However, its limitations may arise when the assumed photometric, noise, data-association, or motion models are violated. ROVIO [96], stereo MSCKF variants [97], factor-recovery mapping [98], ICE-BA [99], robocentric VIO [100], HybVIO [101], and PIVO [102] illustrate different choices in residual construction, state frame, filtering, optimization, mapping, and uncertainty propagation. Related systems such as Kimera [103] and SVO [104] provide relevant spatial-perception and visual-odometry frameworks, respectively, but do not themselves provide direct evidence of performance in lunar VIO applications.

Learning-based VIO estimates pose, motion, or visual–inertial fusion representations primarily through learned mappings. End-to-end methods such as VINet and causal transformer-based VIO [94,95] exemplify this category. These methods may learn complex cross-modal relationships and reduce reliance on manually designed measurement models within the evaluated distributions. However, the available evidence does not establish reliable transfer to representative lunar conditions, calibrated uncertainty, or flight-relevant robustness.

Hybrid VIO architectures integrate learned components with explicit model-based state estimation. Within these architectures, learned modules may support visual feature extraction and matching [19,20,53,54,86,87,88], depth or geometric-prior prediction [89,90,91,92,93], detector-free or dense matching [105,106], semantic landmark detection [21,22,23], and confidence or uncertainty estimation [86,107,108]; shape-from-shading terrain reconstruction in PSRs [109] provides a physics-based, non-learned source of terrain priors. Their outputs may be incorporated as measurements, priors, covariance estimates, or gating scores within a filtering or optimization framework. Learned image-quality and sensor-health assessment are considered separately in Section 5 because evidence of their integration into lunar VIO remains limited [110,111]. Under this taxonomy, a system that uses a learned front end while retaining an explicit geometric or probabilistic estimator is classified as hybrid rather than fully learning-based. In hybrid VIO, the explicit estimator can provide a framework for evaluating learned outputs through geometric residuals, uncertainty calibration, and consistency checks against inertial and auxiliary metric constraints. However, these safeguards require explicit design and validation and do not follow automatically from the adoption of a hybrid architecture.

2.4. Scope Relative to Existing Reviews

The literature relevant to lunar visual–inertial state estimation spans VIO, learning-based simultaneous localization and mapping (SLAM), planetary navigation, and spacecraft fault management. General SLAM and VIO surveys establish the foundations for estimation. Cadena et al. [112] framed robust perception, high-level understanding, and resource awareness as central open problems. Huang [113] reviewed the visual–inertial core, including MSCKF filtering, inertial pre-integration, and optimization-based estimation, while Macario Barros et al. [114] provided a broader visual-SLAM taxonomy. These surveys are platform-agnostic: they explain estimator architecture, but not how lunar photometry, weak excitation, PSRs, and flight SWaP jointly alter its operating envelope.

Learning-based SLAM reviews cover a complementary layer. Chen et al. [84] organized deep localization and mapping into modular versus end-to-end designs, supervision, and uncertainty, and Mokssit et al. [85] extended that account to newer learned front ends. They provide the basis for discussing learned features, depth, matching, and uncertainty, but terrestrial scenes dominate their evidence. They do not establish that these components retain their calibration or robustness on non-Lambertian, high-dynamic-range lunar imagery. Planetary navigation provides stronger operational heritage but only partial environmental transfer. Mars Exploration Rover (MER) visual odometry [71] demonstrates that stereo visual odometry (VO) can operate on flight compute, yet Mars and the Moon differ in illumination, surface photometry, gravity, motion profiles, and available sensing. Terrain-relative navigation reviews and lunar landing demonstrations [5,77] establish map-relative localization during descent, but not sustained post-landing visual–inertial mobility. More recent resources—S3LI, OAISYS, LunarSim, and LuSNAR —more directly expose the domain gap, but are datasets, simulators, or system papers rather than reviews of the complete navigation stack.

Finally, spacecraft AI and fault-management reviews operate mainly above or outside the VIO estimator. Izzo et al. [115] surveyed deep learning across spacecraft guidance, navigation, and control, while fault detection, isolation, and recovery (FDIR) reviews [116] address platform-level health management. They provide useful fault-management context but do not specify how a lunar VIO estimator should use residuals, uncertainty, conditioning, and sensor status to preserve a valid navigation state. The review focus also reflects the spacecraft AI and robotics themes of the target Special Issue [117]. Prior studies provide complementary coverage of general estimation methods, planetary operational experience, learning-based perception, and spacecraft autonomy; however, their integration within a lunar-specific VIO framework remains limited. The remaining gap is therefore a synthesis problem: to determine which techniques address the failure chain in Table 1, what evidence supports their use in lunar transfer, and where that evidence remains simulation-dependent. Table 2 positions the present synthesis against representative reviews and adjacent engineering reports. Four dimensions are discriminating: fused visual–inertial estimation, explicit lunar operating conditions, systematic treatment of learned components, and robust estimator design.

Table 2. Coverage of representative reviews and adjacent engineering studies. VIO = visual–inertial odometry; VI-SLAM = visual–inertial simultaneous localization and mapping; AI = artificial intelligence. Yes = explicit primary objective; Partial = secondary or incomplete coverage; No = absent or incidental background.

Review	Primary Scope	VIO	Lunar	AI	Robustness
Cadena et al. [112]	SLAM robustness	Partial	No	Partial	Partial
Huang [113]	Visual–inertial odometry	Yes	No	No	No
Macario Barros et al. [114]	Visual-SLAM taxonomy	Partial	No	Partial	No
Chen et al. [84]	Deep-learning SLAM	Partial	No	Yes	Partial
Mokssit et al. [85]	Learning-based SLAM	Partial	No	Yes	Partial
Maimone et al. [71]	Mars flight visual odometry	Partial	No	No	No
Johnson and Montgomery [77]	Terrain-relative navigation	Partial	Partial	No	No
Izzo et al. [115]	Spacecraft guidance/navigation/control	No	No	Yes	Partial
This review	Lunar VIO and VI-SLAM	Yes	Yes	Yes	Yes

Decision rules for Table 2: Yes denotes that the topic is an explicit primary objective of the reviewed source and is supported by dedicated methodological development or results. Partial denotes substantive secondary coverage, coverage of only part of the relevant estimation chain, or discussion without end-to-end evaluation. No denotes that the topic is absent or appears only as incidental background without dedicated analysis. These labels describe the scope of coverage, not methodological quality or maturity.

The table is not a claim that every topic is absent from every prior source. It identifies the combination that remains weakly synthesized. The present review uses lunar conditions as the organizing constraint, evaluates AI only where it has a defined function in the navigation pipeline, accounts for scale across multiple sensor configurations, and separates evidence of estimator failure from runtime policies for responding to it.

2.5. Literature Search, Screening, and Evidence Appraisal

This review is a narrative technical synthesis rather than a prospectively registered systematic review; its literature identification and screening are therefore documented retrospectively. Records were identified through targeted searches of IEEE Xplore and Google Scholar, supplemented by forward and backward citation tracing and author-curated sources. The evidence synthesis was limited to the literature available up to 1 August 2026. To document the search and screening process, a retrospective audit was conducted on 25–27 September 2026 while applying the same literature cutoff. No lower publication-year limit was applied. The resulting reference list spans 1981–2026; this range describes the bibliography and was not an eligibility restriction. IEEE Xplore Query B combined “visual-inertial odometry” OR “visual-inertial SLAM” ($n$ = 2361). Within these results, secondary searches for “lunar” and “Moon” yielded seven and 13 records, respectively; all seven “lunar” records were contained within the 13-record “Moon” set. IEEE Xplore Query C combined (“lunar rover” OR “Moon rover”) AND “navigation” ($n$ = 232). Query labels follow the original search log; a planned stand-alone lunar query (Query A) was superseded by the within-Query-B “lunar” and “Moon” searches, so the IEEE Xplore screening flow in Figure 4 is based on these three exports. Google Scholar was searched using “visual-inertial odometry lunar rover” and returned approximately 2250 results, with patents excluded.

Figure 4. PRISMA-ScR-style retrospective documentation of literature identification and screening. The IEEE Xplore branch is a complete retrospective eligibility flow; the Google Scholar, citation-tracing, and author-curation strands are documented only where auditable records were available. The 78 eligible reports are specific to the reconstructed IEEE branch; the 203 references constitute the full cited corpus, not an included-study count.

Across the three documented IEEE Xplore exports, 252 records were collected and 240 unique records remained after removal of 12 duplicates. Title and abstract screening conducted in two passes excluded 151 records, leaving 89 reports for full-text retrieval. All 89 reports were retrieved and assessed in full. Eleven were excluded: four lacked sufficient topical relevance to the review question; three concerned planning, control, platform, or communications without relevant visual-navigation evidence; two addressed powered descent, orbital, or remote sensing applications only; and two addressed non-visual localization only. The remaining 78 met the stated eligibility criteria for the retrospective IEEE Xplore branch. This eligibility count is specific to the reconstructed IEEE branch and should not be interpreted as the number of studies incorporated into the manuscript’s narrative synthesis. Google Scholar screening was limited to the first 100 relevance-ranked results, of which approximately 63 were judged topic-relevant and approximately 37 were excluded. Because no item-level export was retained for the Google Scholar strand, subsequent duplicate-removal and full-text counts could not be reconstructed. Counts for historical citation tracing and author-curated sources were likewise not retrospectively recorded. Figure 4 therefore presents a complete retrospective eligibility flow for the IEEE Xplore branch and transparent partial documentation for the other, overlapping strands; these strands are not summed.

Eligible sources addressed visual–inertial estimation, lunar-relevant perception, metric-scale constraints, robust estimation, failure handling, datasets, or onboard implementation. Studies limited to powered descent, orbital rendezvous, pure remote sensing, or non-visual localization without a relevant visual component were excluded. Screening was performed by the first author without independent duplicate assessment. Eligibility screening was separate from evidence appraisal. Given the heterogeneity of the reviewed evidence, no common numerical quality or maturity score was assigned; instead, each source was characterized by the type of claim its evidence can support, using the decision rules given with Table 3 and Table 4. The reference list contains 203 contextual, dataset, technical, and author-curated sources. It should not be interpreted as either 203 included studies or as a one-to-one list of the 78 reports meeting the retrospective IEEE eligibility criteria.

3. AI-Enhanced Front Ends for Lunar VIO

This section examines whether and under what conditions model-based, learning-based, and hybrid visual front ends improve the availability and reliability of measurements supplied to a VIO estimator. Learned local features and matching, semantic landmarks, geometric priors, and illumination-aware processing are assessed with respect to lunar operating conditions, the validation domain of the available evidence, and reported computational requirements. Handcrafted feature and direct photometric methods are retained as reference baselines.

3.1. Lunar Visual Degeneracy

In the visual–inertial estimation literature, the term visual degeneracy is used at two levels. At the estimator level, degeneracy denotes rank deficiency in the visual contribution to the information matrix, whereas near-degeneracy is reflected in poor numerical conditioning. A commonly used operational characterization examines the eigenvalues of the visual information matrix, particularly its smallest eigenvalue [38]. At the front-end level—which is what concerns us in this section—visual degeneracy denotes the more concrete phenomenon of the front-end producing measurements that are either too few, too poorly localized, or too geometrically degenerate to constrain the camera’s motion. The same information matrix links the two levels. For residual Jacobians ${\mathit{J}}_i$ with covariance ${\mathsf{\Sigma }}_i$, the local matrix $\mathsf{\Lambda }=\sum _i{\mathit{J}}_i^{\mathrm{\top }}{\mathsf{\Sigma }}_i^{-1}{\mathit{J}}_i$ summarizes finite-horizon conditioning: rank deficiency or a small minimum eigenvalue indicates weakly constrained directions for the current trajectory and measurement geometry [38]. Front-end measurement deficiencies can contribute to weak estimator conditioning, although their effects on the estimated state require estimator-level information or observability analysis.

The literature suggests three recurring, but configuration-dependent, measurement risks. Low-texture or repetitive terrain can reduce the number and spatial distribution of reliable correspondences. Hard illumination transitions and view-dependent reflectance may degrade brightness-based tracking or produce unstable gradients [13,14]. Limited depth variation and certain motion geometries may weaken directional constraints [118]. Their prevalence depends on terrain, illumination, camera configuration, exposure, and trajectory; the available evidence does not justify treating any one of them as universal across the Moon. Polar illumination models indicate that local illumination depends on location, topography, solar geometry, and secondary reflection [119,120,121,122,123]; benchmark design should therefore represent variation within permanently shadowed regions (PSRs), rather than treating all PSRs as a single, uniformly dark condition [109,121,122,123].

3.2. Learning-Based Local Features and Matching

One learning-based design replaces or augments hand-crafted detectors and descriptors with learned alternatives. The relevant question is not terrestrial benchmark rank alone, but whether match repeatability, spatial coverage, calibration, and runtime transfer to representative lunar data. Representative learned features include SuperPoint [19], reliability-aware R2D2 [86], policy-trained DISK [87], deformable ALIKED [88], and efficiency-oriented XFeat [20]. They illustrate different trade-offs among repeatability, ambiguity rejection, viewpoint tolerance, and compute, but their reported terrestrial results are not evidence of lunar superiority. D2-Net, LoFTR, DKM, and RoMa broaden the comparison from sparse learned detectors to detector-free and dense matching, but their lunar calibration and computational trade-offs remain unestablished [105,106,124,125].

Direct lunar evaluations of these descriptors remain sparse; most traceable evidence comes from simulator and planetary-analog resources rather than lunar flight telemetry [59,61,62]. LuVo uses LightGlue inside a stereo lunar–VO pipeline [54], while BEVIO studies feature association under sparse day/night updates in simulation and on a rover analog [53]. These studies support the value of learned matching or geometry-aware warping in specific conditions, but they do not establish uniform superiority across mare, highlands, sunlit terrain, and deep shadow. A controlled cross-descriptor lunar benchmark remains missing. More numerous and better distributed correct matches can improve the finite-horizon visual information available to the back end. A learned descriptor cannot, however, create geometric information absent from the scene, and current lunar comparisons do not establish a general conditioning improvement. Figure 5 contrasts the measurement patterns produced by representative point-tracking and matching strategies.

Figure 5. Schematic comparison of visual tracking and matching on a terrestrial indoor corridor sequence. The corridor is used as an illustrative example because, like much lunar terrain, it is dominated by repetitive structure and weak texture; it does not reproduce lunar illumination or surface reflectance. (a) Pyramidal optical flow propagates selected points between nearby frames but may lose tracks or produce outliers at strong shadows; (b) Oriented FAST and Rotated BRIEF (ORB)—where FAST means Features from Accelerated Segment Test and BRIEF means Binary Robust Independent Elementary Features—detects oriented keypoints and computes binary descriptors for frame-to-frame association; (c) learned or geometry-aware matchers establish wider-baseline correspondences and require explicit outlier filtering. The overlays illustrate algorithm behavior and are not comparative benchmark results; (d) actual neural energy activation illustration from the learning-based feature description.

3.3. Learning-Based Semantic Lunar Landmarks

Semantic detectors can supplement generic keypoints with lunar-relevant structures that may also carry catalog or geometric priors. Their contribution is conditional on detectability, correct association, and the compatibility of orbital and rover viewpoints. Crater detection: A prominent lunar landmark class. Impact craters are widespread and can provide recognizable geometric structure and catalog-linked priors. Conic projections can often approximate their rims, and their diameters and center coordinates may be cross-referenced against orbital catalogs. Detectability from a rover is nevertheless conditional on terrain, range, viewpoint, illumination, image quality, and catalog coverage. Two broad generations of crater detectors are represented in the literature. Recent object-oriented and foundation-model methods extend this direction by targeting explicit lunar SLAM and zero-shot geological landmark matching [55,126].

Classical template and edge-fitting approaches [21,22] and later convolutional detectors [23,24,25,26] have mainly been evaluated on orbital imagery or DEMs. Those results support crater detection in their tested domains but do not establish equivalent performance for oblique, near-field rover views. When a rover-view crater is correctly associated with a catalog entry, known geometric and map information can add an absolute constraint to the estimator [49,50,51]. This action is the principal link between semantic perception and metric state estimation. Its effectiveness depends on visibility, catalog coverage, association accuracy, and uncertainty modeling. Flight and high-fidelity crater-navigation studies also provide stronger geometric and association baselines than orbital detector accuracy alone [127,128,129,130,131]. General onboard optical-navigation image-processing foundations provide additional context for these landmark pipelines [132]. Global lunar crater catalogs provide candidate priors, although the catalog scale and completeness must match the rover’s operating range [133].

Boulders and isolated rocks are plausible local semantic landmarks, but the cited orbital mapping evidence [134] does not establish a validated boulder-based lunar VI-SLAM system. They are therefore retained only as a prospective option. ShadowNav combines enhancement, stereo geometry, active illumination, and crater-edge matching [49], demonstrating one specific localization configuration rather than a general shadow-mask solution for VIO. Exposure modeling and illumination-robust methods [135,136] provide transferable components; their activation and weighting should be treated as estimator configuration and robustness questions.

3.4. Learning-Based Depth and Geometric Priors

A third family of learning-based front-end designs supplies the back end with learned geometric priors—typically per-pixel depth, but also surface normals, planar segmentation, or relative pose between consecutive frames—that supplement the sparse keypoint stream. The proximate motivation, especially for monocular configurations, is to recover information about scene depth that the sparse front end leaves implicit and that the back end would otherwise have to estimate from inertial excitation alone. Modern models range from relative-depth methods such as MiDaS [89] to domain-conditioned or large-scale models including ZoeDepth and Depth Anything [90,91,92]. Their terrestrial and zero-shot results motivate evaluation as soft geometric priors, but do not validate metric lunar depth. In a VIO back end, learned depth is mostly treated as a soft residual with explicit uncertainty rather than as an unquestioned metric measurement. Self-supervised depth and transformer-based dense prediction supply additional candidate priors whose metric validity must still be tested under lunar domain shift [137,138].

Most cited networks are trained predominantly on terrestrial imagery. LuSNAR supplies simulated lunar depth and semantics [62], but the review identified no controlled study establishing transfer error across representative lunar terrain and illumination classes. Uncalibrated fusion may therefore introduce systematic error. A metric output learned from terrestrial training data is not, by itself, a physical lunar scale anchor. Until validated, it should be fused as a relative or uncertain prior whose scale is constrained by IMU excitation, stereo, or map-linked geometry. Learned surface normals [93] and map-derived DEM normals [52] are distinct priors. The former remain prospective for lunar VIO; the latter inherit map resolution and registration errors. End-to-end learned VIO maps images and IMU samples directly to motion [94,95]. Its lunar out-of-distribution behavior and uncertainty have not been established. The reviewed evidence therefore supports discussing it as an experimental prior or comparison, not ruling it in or out as a flight estimator. DROID-SLAM, DeepVIO, SelfVIO, and TLIO represent recurrent optimization, self-supervised visual–inertial fusion, and learned inertial fallback, respectively [139,140,141,142].

3.5. Lunar Direct Methods and AI Computational Constraints

Direct methods such as Large-Scale Direct Monocular Simultaneous Localization and Mapping (LSD-SLAM), Direct Sparse Odometry (DSO), and Delayed Marginalization Visual–Inertial Odometry (DM-VIO) [143,144,145] rely on photometric residuals whose standard models can be stressed by view-dependent reflectance, clipping, exposure changes, and albedo variation. Calibration and illumination-aware models [135,136] may mitigate these effects. Because controlled lunar comparisons are lacking, the defensible conclusion is that risk is configuration-dependent rather than categorical inferiority to feature-based methods. Planetary-mission testing also provides a recent robust monocular VO comparison under visually difficult analog conditions [146]. An adequate comparison should therefore vary exposure response, shadow fraction, terrain texture, and viewpoint while holding the estimator and compute budget fixed. Reporting only trajectory error would hide whether failure originated in photometric calibration, correspondence loss, or back-end conditioning; front-end survival rate, inlier spatial coverage, and residual stability should be reported alongside pose error.

Deployment depends on resolution, frame rate, precision, hardware, and duty cycle as well as accuracy. Lightweight features such as XFeat [20] and larger depth models [92] occupy different compute regimes, but operation counts alone do not establish flight feasibility [17,18]. Quantization, distillation, and hardware co-design [147,148,149] are candidate implementation tools. Their lunar accuracy–resource trade-offs must be measured on a specified target platform rather than inferred from terrestrial benchmarks. Field-programmable gate array (FPGA) rover visual odometry and the Navion accelerator demonstrate that major reductions in latency and power are feasible, although neither result alone establishes lunar flight qualification [150,151]. For flight-oriented studies, latency distributions, peak memory, energy per update, thermal throttling, and recovery after a missed deadline are as important as mean throughput. Measurements should also include the geometric verification and uncertainty calibration stages surrounding the neural model, because an apparently lightweight network can shift rather than reduce system cost.

3.6. Lunar Front-End and AI Findings

The front-end evidence supports three bounded conclusions. First, learned features, semantic landmarks, illumination handling, and geometric priors address different measurement risks; none has been shown to dominate across representative lunar conditions. Second, crater-linked constraints provide a clear bridge from perception to metric estimation, but depend on viewpoint, catalog quality, and association quality. Third, direct lunar comparisons and target-hardware measurements remain too sparse for universal accuracy or SWaP rankings. Reliable measurements are necessary but do not ensure a metrically trustworthy state. The next section therefore evaluates the back-end property that front-end performance alone cannot establish: whether visual, inertial, stereo, wheel, and map constraints produce a stable and compatible metric state. This evidence suggests a practical experimental design that crosses terrain and illumination strata with sensor configuration, front-end family, and compute target. Such a design would show whether a gain comes from better measurement availability, improved outlier rejection, or merely a more favorable operating point, and would make negative or conditional results reusable across missions. The metrics required at each evaluation level are consolidated in the recommended evaluation matrix in Section 6.3.

4. Lunar VIO Estimation, AI Integration, and Metric-Scale Integrity

This section examines the tightly coupled VIO estimator: whether the metric scale is identifiable, sufficiently conditioned, and accurate over the estimation horizon, and whether it is continuous across time and changes in the metric source. Monocular VIO is the limiting theoretical case; stereo, wheel, crater, and map constraints add metric information but introduce configuration-specific calibration, association, registration, or slip risks. The analysis separates structural observability from practical conditioning, then compares the metric information and failure modes of representative sensor configurations. Within this lunar estimator, AI is relevant when it supplies visual correspondences, semantic or depth constraints, heteroscedastic uncertainty, source-quality scores, or health indicators. These outputs can improve the availability of practical information, but they do not change the structural observability of the underlying camera–IMU model.

4.1. Lunar Estimation and AI Evaluation Criteria

Metric-scale integrity is used here as an operational umbrella rather than as a standardized term from a single prior theory. It comprises four related properties. Observability asks whether the metric scale is identifiable from the sensor model and the trajectory under the stated assumptions. Conditioning describes the sensitivity of that estimate to finite-horizon geometry, noise, bias, and calibration. Metric accuracy concerns agreement with physical distance and velocity. Temporal and cross-source consistency concerns continuity over initialization, keyframe and map updates, loop closure, reinitialization, and sensor handover. This last meaning must not be conflated with statistical estimator consistency, which compares reported uncertainty with realized error. An AI component used in lunar VIO must additionally be evaluated for calibration, out-of-distribution behavior, geometric consistency, temporal stability, false confidence, and compute cost [84,85,107,108]. Formally, observability and conditioning are analyzed through the visual–inertial information matrix $\Lambda$ (Section 3.1) and its minimum eigenvalue (Section 5.2), together with the accelerometer scale-recovery model and excitation–bias ratio in Section 4.3 (Equations (5) and (6)), while metric accuracy and cross-source consistency concern the agreement of the local estimate with ground truth and with the auxiliary metric constraints in Equation (4). For evaluation, metric accuracy over a traverse segment can be summarized by the scale error $e_s=\left|‖{\widehat{P}}_k-{\widehat{P}}_0‖/‖P_k-P_0‖-1\right|$, where $\widehat{P}$ and $P$ denote estimated and ground-truth positions, and cross-source consistency by the normalized discontinuity $d=\sqrt{{\left({\widehat{x}}^{+}\boxminus {\widehat{x}}^{-}\right)}^{\top }{\Sigma }^{-1}\left({\widehat{x}}^{+}\boxminus {\widehat{x}}^{-}\right)}$ at a reinitialization, map update, or source handover, where ${\widehat{x}}^{-}$ and ${\widehat{x}}^{+}$ are the state estimates immediately before and after the event, $\boxminus$ denotes the difference on the state manifold, and $\Sigma$ is the associated covariance.

Metric-scale error matters because downstream consumers express motion, hazards, and map constraints in physical units. This review does not analyze planning or control; it uses those interfaces only to explain why the estimator must report a metrically meaningful state and uncertainty. Wheel-slip checks and map fusion likewise require compatible local and auxiliary metric estimates [112,152,153]. Figure 6 summarizes the fixed-lag factor graph underlying the optimization-based estimators discussed below.

Figure 6. Lunar VIO optimization back end with bounded AI components. IMU pre-integration links consecutive navigation states; geometric visual factors connect states to landmarks; stereo, wheel, crater, and DEM constraints provide configuration-dependent lunar metric information. Learned features, semantic landmarks, depth priors, or confidence may enter as validated factors with explicit uncertainty. Robust fixed-lag optimization with marginalization retains responsibility for motion, bias, scale-related quantities, calibration, consistency, and marginalized prior information.

4.2. Observability of Monocular Lunar VIO

Classical monocular visual–inertial observability results provide the baseline against which lunar operating conditions must be interpreted. Under standard assumptions, monocular VI-SLAM retains four gauge freedoms: global position and rotation about gravity [154,155]. These structural freedoms are distinct from metric scale and must be handled consistently by the estimator. Consistency-preserving extended Kalman filter (EKF) linearization, observability-constrained design, heterogeneous geometric features, and invariant filtering provide complementary formulations for preserving the correct unobservable subspace [40,41,42,43]. Metric scale is conditionally observable under sufficiently informative visual–inertial motion [39]. The result depends on excitation that separates translational acceleration, gravity, and bias over the estimation horizon; it does not mean that gravity magnitude alone determines scale quality. Specific motion classes can introduce additional unobservable or weakly conditioned directions that couple scale and bias [155,156]. Some smooth, nearly constant-motion rover segments may approach such conditions, but this is trajectory-dependent and should not be generalized to lunar mobility as a whole. Accelerometer and gyroscope biases are also observable only under appropriate motion [154,157]; weak bias estimation can in turn degrade scale conditioning. Online temporal calibration and dedicated initialization methods further show that timing, gravity, velocity, scale, and bias should be analyzed jointly rather than treated as independent setup steps [44,45,46]. Dedicated VIO-initialization benchmarks provide an additional basis for evaluating convergence speed and the accuracy of scale, gravity, velocity, and bias estimates [158]. Existing sufficient conditions do not directly provide mission-specific bounds linking motion, sensor noise, geometry, horizon length, and scale error. That translation from theory to quantitative design requirements remains an evidence gap. Learned features or depth can improve measurement density in a tested lunar regime, but no learned representation can remove gauge freedom or recover information absent from the sensor trajectory; AI benefits must therefore be reported as conditioning, accuracy, coverage, or failure-rate improvements rather than altered observability.

4.3. Effects of Lunar Gravity and Motion

Building on the analysis developed in Section 4.2, we can now characterize precisely what changes when the underlying gravitational acceleration is $g_{Moon}\approx 1.625\hspace{0.17em}{\mathrm{m}/\mathrm{s}}^2$ rather than $g_{\oplus }\approx 9.81\hspace{0.17em}{\mathrm{m}/\mathrm{s}}^2$. The accelerometer signal serves as the scale-recovery channel. The accelerometer measurement model, written in the body frame and ignoring rotation rate effects within a short integration step, takes the standard form:

$${\mathbf{a}}_m\left(t\right)={\mathit{R}}_W^t\left(\ddot{\mathit{p}}\left(t\right)+\mathbf{g}\mathrm{ᵂ}\right)+{\mathbf{b}}_{a_t}+{\mathbf{n}}^a,$$

(6)

where ${\mathit{R}}_W^t\in \mathrm{S}\mathrm{O}\left(3\right)$ is the rotation matrix from the world frame to the body frame at time t, $\ddot{\mathit{p}}\left(t\right)$ is the world-frame linear acceleration, $\mathbf{g}\mathrm{ᵂ}$ is the gravity term defined after Equation (3), ${\mathbf{b}}_{a_t}$ is the slowly varying accelerometer bias, and ${\mathbf{n}}^a$ is the white noise. Metric scale can be recovered when vision constrains the direction of motion (up to scale), and the visual–inertial fusion problem must therefore explain the residual signal ${\mathit{R}}_W^t\ddot{\mathit{p}}$ from the accelerometer, in turn requiring that this signal be distinguishable from the dominant gravitational contribution ${\mathit{R}}_W^t\mathbf{g}\mathrm{ᵂ}$. In a level traverse with bounded translational acceleration, a potential diagnostic is the ratio between the non-gravitational excitation and the accelerometer-bias uncertainty:

$${\rho }_{\mathrm{e}\mathrm{x}\mathrm{c}}={\displaystyle \frac{‖\ddot{\mathit{p}}‖}{{\sigma }_{b_a}}},$$

(7)

where ${\sigma }_{b_a}$ is expressed in acceleration units representing the uncertainty of the accelerometer bias. This process is a conditioning heuristic, not a formal observability theorem: values near unity indicate that translational excitation can be difficult to distinguish from bias. Lunar gravity changes the specific-force decomposition and attitude reference, but low $\left|\mathit{g}\right|$ alone does not imply poorer scale recovery. The practical concern is the joint regime of slow, smooth rover motion and constrained inertial sensing. Ref. [71] documents Mars rover traverses in such a low-excitation regime, but this does not establish a universal lunar-to-Earth penalty. Mission-specific motion profiles and IMU specifications are therefore required to quantify conditioning. In that regime, the conditional observability established by [39] can remain formally intact while being practically weak. Temperature-dependent IMU bias may also affect long traverses [17,159,160], but its magnitude is sensor- and thermal-design-specific. It should be treated as a calibration and validation requirement rather than a universal failure of the lunar scale. Reduced gravity changes the specific force and attitude reference environment but, by itself, does not make the metric scale unobservable. The supported conclusion is conditional: low excitation, finite-grade inertial sensing, and weak visual geometry can jointly produce poor finite-horizon scale–bias conditioning.

4.4. Lunar Sensor, AI, and Compute Constraints

SWaP affects practical metric integrity through sensor grade, measurement rate, image quality, and feasible estimation horizon; it does not, by itself, change structural observability. Inertial sensors exhibit broad trade-offs among stability, power, mass, and thermal design [159,160], but product-level variation prevents a universal grade-to-performance mapping. Lunar system claims should therefore use mission-specific IMU specifications and temperature calibration rather than a generic, approximate taxonomy. The resulting design variable is not sensor-grade alone, but the combined error budget across thermal calibration, sampling, synchronization, and the estimator horizon. Optical dynamic range, frame rate, resolution, and field of view jointly determine the rate and geometry of visual constraints, as well as the associated compute load. High dynamic range (HDR), event-based, or scheduled exposure can help in selected regimes but carry configuration-specific costs. BEVIO shows that sparse visual updates can work in its tested architecture [53], not that one update rate is generally sufficient. Front-end, back-end, and health-monitor workloads must be evaluated together for processing time, memory, and power on the intended flight processor [17,18]. Learning-based front ends add model storage, memory traffic, accelerator dependence, and confidence-calibration workloads; quantization, distillation, and duty cycling can reduce cost but require lunar-regime accuracy and uncertainty checks [147,148].

The aggregate effect of these three axes is configuration-dependent. An IMU, camera, and compute platform selected within a tight SWaP budget may provide less favorable scale–bias conditioning than a higher-resource terrestrial system, particularly during low-excitation motion. SWaP does not alter structural observability in itself; it constrains sensor quality, measurement rate, and estimation horizon and can therefore weaken practical metric-scale integrity even when scale remains theoretically observable.

4.5. Lunar Metric Constraints and AI-Assisted Consistency

Metric sources are compared based on the information they provide and the mechanisms by which that information can become biased, intermittent, or incompatible with the current state. Calibrated stereo supplies direct metric depth when correspondence and disparity are adequate, reducing reliance on inertial excitation [12,54,161]. Distant or low-texture scenes, calibration drift, thermal baseline changes, and monocular fallback can still weaken or interrupt that cue. Its suitability therefore depends on geometry and platform cost, not baseline alone.

A correctly associated, catalog-linked crater can provide an absolute map and geometric constraint [49,50,51]. Availability, catalog coverage, oblique-view detection, and false association limit when that constraint is valid; an accepted mismatch can create an abrupt state correction. DEM or orbital-map matching can provide periodic absolute constraints [52,77], but accuracy and consistency depend on map resolution, datum, registration, and terrain-match uncertainty. It is therefore a complementary global cue rather than a substitute for local metric estimation. Rover-to-orbital localization has been demonstrated with ground imagery, boulder registration, LiDAR, and onboard global-localization systems, but each method inherits map-resolution and registration assumptions [78,79,80,81,82,83].

Wheel odometry provides a low-power metric distance cue for wheeled platforms, while slip and sinkage can bias it [16,152,153]. Its contribution should therefore be weighted by contact or innovation evidence and does not transfer to legged or hopping configurations. Learned monocular depth can enter as a soft prior [91,162], but no lunar-specific fusion evidence identified here establishes its metric calibration or temporal consistency. A terrestrial metric-depth output should not be treated as an absolute lunar anchor without validation. Gravity-direction information is already part of visual–inertial estimation [39,155]. Additional lunar gravity priors may improve attitude or conditioning in specified models, but the review found limited lunar evaluation and does not claim a direct independent scale source. Current-based slip detection, visual–inertial slip compensation, and terrain-conditioned slip prediction offer complementary approaches for identifying or correcting corrupted wheel constraints [163,164,165,166]. Learned confidence or terrain classification may modulate these auxiliary factors, but the resulting weights must be calibrated against lunar or representative analog data and checked against estimator innovations rather than accepted as absolute truth [107,108,163,164,165,166]. The reviewed systems combine local odometry with selected geometric, map, illumination, or inertial cues [49,51,52,53], but no common architecture implements and evaluates handover among all sources. Metric compatibility at reinitialization, map update, or sensor handover is therefore a core estimator property and a principal input to failure detection and recovery.

4.6. Lunar Estimator and AI Integration Findings

The estimator analysis yields three conclusions. Scale is conditionally observable in monocular visual–inertial estimation, but practical accuracy depends jointly on motion, bias, geometry, and horizon. Nominally metric stereo, wheel, crater, and map cues remain vulnerable to distinct consistency failures. Finally, current studies lack a shared protocol for unaligned scale error, temporal drift, and discontinuities at reinitialization or handover. Quantitative lunar-specific bounds and comparable datasets remain priorities rather than established results. A fourth conclusion concerns AI: learned measurements and confidence can strengthen finite-horizon lunar VIO only when their calibration, domain validity, geometric compatibility, uncertainty, and fallback behavior are explicitly represented. Metric-scale integrity is therefore configuration-dependent. Its observable indicators—visual conditioning, residuals, uncertainty, calibration checks, and disagreement among metric sources—provide the evidence needed for robust estimation and failure handling.

5. Robust Lunar VIO and AI-Assisted Failure Recovery

This section examines how a VIO estimator detects unreliable measurements, weak state conditioning, sensor faults, and resource-induced degradation, while preserving a valid navigation state. Model-based mechanisms form the evidence baseline; learned components are included only where they perform a defined estimation, confidence, or health function. The visual front end provides candidate measurements, while the tightly coupled back end defines the metric state and the uncertainty to be preserved. Robust estimation connects these layers through measurement rejection, reweighting, constrained updates, source fallback, reinitialization, and explicit health reporting. Learning-based fault or confidence classifiers are included only when their outputs are calibrated, monitored for lunar domain shift, and coupled to a bounded estimator response.

5.1. Robustness Requirements

The resulting supervisory layer may gate or reweight measurements, select a fallback source, alter window or update settings, and report health to higher-level autonomy. These functions have different levels of evidence maturity and should not be presented as a single established autonomous-agent architecture. Three bounded requirements motivate this layer: measurement quality varies across a traverse; calibration, bias, and source compatibility can degrade; and higher-level autonomy requires a trustworthy state, uncertainty, and health output. The objective is not general decision-making but preservation of estimator validity. Operationally, these requirements imply more than detecting a bad frame. The estimator must distinguish transient visual degradation from persistent calibration or sensor faults, prevent a low-quality source from corrupting other factors, and expose whether the current state is nominal, degraded, or invalid. Each health transition should be tied to observable evidence and a bounded response, with hysteresis or dwell logic to prevent rapid switching that could create discontinuities. Validation should measure detection delay, false alarms, accumulated state error during degradation, and recovery transients, rather than reporting nominal trajectory accuracy alone.

5.2. Degeneracy Detection, Constrained Updates, and Window Management

The most robust function is the detection of front-end or back-end degeneracy, followed by a controlled change in the state update. Hand-crafted degeneracy detectors: The seminal contribution is [38], which analyzes degeneracy via the eigendecomposition of the optimization Hessian or pseudo-Hessian. A small minimum eigenvalue indicates weak finite-horizon information, while the associated eigenvector identifies the corresponding weakly constrained state direction. A zero eigenvalue under the stated linearization indicates rank deficiency; a merely small value should not be described as proof of strict unobservability. Candidate responses include projecting, damping, or withholding updates along weak directions, depending on the estimator and its consistency requirements [38,47,48]. Condition-number and per-direction indicators follow from the same information-matrix analysis [155], but their thresholds remain sensor-, state-, horizon-, and operating-regime-dependent.

Analytic degeneracy indicators require thresholds and modeling choices whose usefulness depends on sensor noise, state parameterization, the horizon, and the operating regime. This move motivates adaptive or learned thresholds as a hypothesis, not evidence that fixed thresholds necessarily fail on the Moon. A learned classifier could use match distribution, residuals, or visual–inertial agreement, but the reviewed evidence remains dominated by analytic measures [38,47,48]. Terrestrial domain adaptation work [167] motivates the experiments; it does not demonstrate lunar directional-degeneracy classification. Detection is useful only when linked to a defined response. Supported patterns include projecting or withholding updates along weak directions [38] and temporarily relying on inertial, wheel, stereo, or map prediction when a visual measurement stream is unavailable. Existing lunar systems expose complementary cues [49,51,52,53,54], but do not establish a common switching architecture or prove that switching the entire back-end formulation is beneficial. Sliding-window estimators bound computation through keyframe selection and marginalization [11,168]. Information-aware feature or graph sparsification [169,170] and delayed marginalization [145] illustrate accuracy–resource trade-offs. No cited study establishes a lunar-specific adaptive window mechanism that is jointly validated against estimator accuracy, numerical conditioning, runtime, and power.

5.3. Outlier Rejection and Uncertainty

Measurement acceptance ranges from model-based sampling and robust costs to learned match scores and uncertainty. The key boundary is whether a method merely outputs a score or demonstrates calibrated-estimator behavior. Random sample consensus (RANSAC)-family methods [171,172] provide a model-based baseline. Learned correspondence models [173,174,175] can produce scores and cleaner match sets; LuVo supplies lunar-simulation evidence for learned matching [54], but does not establish calibration or transfer across lunar terrain.

Robust losses and graduated non-convexity [176,177] provide back-end alternatives for limiting the influence of outliers. Their relevance to lunar high-outlier regimes is plausible, but comparative lunar evidence was not identified. Learned covariance or confidence is useful only if empirical coverage is calibrated, especially under distribution shift [107,108]. Lunar-specific demonstrations remain sparse, so learned uncertainty is treated as an input requiring validation rather than as an established robustness mechanism. Dynamic covariance scaling, switchable constraints, Max-Mixtures, TEASER, adaptive robust losses, MAGSAC++, and riSAM provide established alternatives for downweighting, modeling, certifying, or incrementally rejecting outliers [178,179,180,181,182,183,184]. Calibration should be assessed using reliability diagrams or empirical coverage across separate terrain and illumination strata, not just by ranking matches within a single sequence. The downstream test is whether the reported uncertainty predicts residual and trajectory error sufficiently well to support gating; an overconfident learned score can be more damaging than a conservative but less accurate one.

5.4. Lunar Sensor Health, AI-Assisted Fusion, and Resource Configuration

Sensor-health monitoring lies at the estimator boundary: it detects whether camera or IMU measurements remain usable and passes a health indication upward. Mission-level recovery actions remain outside the scope of this review. Potential camera degradations include saturation, occlusion, or dust contamination, and temperature-dependent focus changes [185]. Their occurrence and severity are design- and mission-specific. Generic image-quality models [110] offer possible detectors, but the review identified no integrated lunar VIO validation. Bias jumps, saturation, and camera–IMU timing errors can invalidate nominal bias models. Innovation tests and statistical gating provide established tools [111], while performance during long stationary periods and lunar thermal operation requires configuration-specific validation. Multiple-model filters [186] and configurable frameworks such as OpenVINS [9] provide relevant architectural elements, but the cited sources do not establish a lunar health-driven soft-switching system. Discrete and continuous weighting therefore remain alternatives for evaluating the estimator—health interface. The VIO boundary adopted here ends with a state estimate, its uncertainty, and a health indication passed to higher-level autonomy. Halting, replanning, and vehicle-level reconfiguration remain outside the scope of the review.

Resource-aware estimator configuration is comparatively immature for lunar VIO. Window size, module activation, image resolution, and sensor update rate influence both estimator quality and SWaP. Existing work provides information-aware selection and active-perception concepts [169,170,187], but no direct lunar VIO study establishes a validated controller that jointly guarantees estimation quality and resource limits. Reinforcement learning [188] may be considered an implementation option, but the VIO architecture does not require it and is not treated as an established navigation component. Figure 7 integrates these tests and bounded responses into an estimator-level health-and-recovery loop.

Figure 7. Estimator-level robustness and failure-recovery loop for lunar VIO. Measurement-quality tests, finite-horizon conditioning, source agreement, timing, sensor status, and resource pressure inform a model-based or learning-based health state. Bounded responses include gating or reweighting measurements, constraining weak states, resizing the estimation window, switching available metric sources, and reinitialization, with the resulting state and uncertainty reported as nominal, degraded, or invalid. The loop closes through the feedback path from the bounded responses back to the estimator: after any bounded response, the estimator continues, and its updated measurements and state re-enter the same tests at the next epoch. Mission planning and rover control are outside the scope of this review.

5.5. Robustness Findings

Robust lunar VIO is currently supported by a collection of mechanisms rather than a single validated architecture. Analytic gating, robust estimation, innovation testing, and bounded-window estimation have the strongest foundations. Lunar evidence remains limited for learned degeneracy classification, calibrated learned uncertainty, integrated sensor-health fallback, and adaptive resource configuration. The supported near-term design is therefore an explicit model-based estimator with measurable health indicators and carefully validated learned components. A credible integrated evaluation should therefore inject defined visual, inertial, timing, and compute faults and record detection delay, false-alarm rate, peak state error, uncertainty coverage, and recovery continuity. This action would separate mechanisms that improve average accuracy from those that preserve estimator validity during rare but consequential failures.

6. Lunar VIO Systems, AI Components, Datasets, and Evaluation

Section 3, Section 4 and Section 5 analyzed perception, metric-scale integrity, and estimator-level responses separately. This section recombines the evidence at two levels. Section 6.1 compares representative systems and the evidence boundaries of their reported mechanisms; Section 6.2 examines the datasets and facilities that support—or limit—sim-to-real claims; and Section 6.3 draws the joint implications. Architecture and evaluation evidence are kept distinct because a plausible system design is not flight-relevant merely because it performs well in one simulator or analog. For AI-enhanced systems, the evaluation must separate gains from the learned component, the geometric estimator, and auxiliary lunar metric sources, and must report domain, training data, calibration, and target-hardware cost.

6.1. Representative Methods, Demonstrations, and Evidence for Lunar VIO

Table 3 is a selective comparison of terrestrial reference methods, operational planetary localization evidence, lunar surface-mobility prototypes, and end-to-end learned VIO. The terrestrial block is limited to exemplars needed to distinguish filtering, keyframe and direct optimization, platform, multi-map, and probabilistic formulations; additional terrestrial variants are discussed in Section 2.3 [96,101,103]. MER visual odometry [71] and Yutu localization [16] provide operational evidence, whereas crater-based methods [49,50,51], LuVo [54], DEM-anchored SLAM [52], BEVIO [53], and Lo-SLAM [55] remain prototypes or studies at the evidence levels stated in the final column. For compactness, Table 3 combines role with estimator, inputs with the source of metric information, and any learned component with the principal distinguishing mechanism. Labels follow what each cited source supports: a method paper, prototype, analog demonstration, or operational result is not promoted to a flight system without corresponding evidence. In particular, PIVO [102] is a probabilistic inertial–visual odometry method and is classified as a terrestrial reference method, not as a lunar or planetary operational system. Seeker [76] is retained in the surrounding discussion as an integrated Earth-analog rover demonstration rather than a directly comparable VIO estimator.

Table 3. Representative methods, demonstrations, and evidence relevant to lunar VIO. Evidence labels state the reported validation boundary and do not imply equal maturity or flight readiness.

Representative Work (Year)	Role/Estimator	Inputs/Metric Source	Learned Component or Distinguishing Mechanism	Reported Evidence Boundary
Terrestrial reference methods (not lunar or planetary systems).
MSCKF [8] (2007)	Filter-based VIO (MSCKF)	Monocular + IMU/inertial scale	No learned component; multi-state feature constraints	Method paper; terrestrial tests; no lunar evidence
OKVIS [168] (2015)	Keyframe VIO; nonlinear optimization	Stereo + IMU/stereo–inertial scale	No learned component; tightly coupled optimization	Method paper; terrestrial tests; no lunar evidence
VINS-Mono/VINS-Fusion [11,161] (2018/2019)	Optimization-based VIO/multi-sensor odometry	Monocular or stereo + IMU/inertial or stereo scale	No learned component; multi-sensor optimization	Method/framework papers; terrestrial tests
OpenVINS [9] (2020)	Research platform; MSCKF family	Monocular or stereo + IMU/inertial or stereo scale	No learned component; configurable estimator platform	Platform paper; terrestrial tests; no lunar evidence
ORB-SLAM3 [12] (2021)	Visual–inertial SLAM; full optimization	Monocular or stereo + IMU/inertial or stereo scale	No learned component; multi-map SLAM	Method paper; terrestrial tests; no lunar evidence
DM-VIO [145] (2022)	Direct VIO; nonlinear optimization	Monocular + IMU/inertial scale	No learned component; delayed marginalization	Method paper; terrestrial tests; no lunar evidence
PIVO [102] (2018)	Probabilistic inertial–visual odometry (EKF)	Monocular + IMU/inertial scale	No learned component; probabilistic occlusion handling	Terrestrial test (WACV method paper); no planetary or flight evidence
Operational planetary localization evidence.
MER visual odometry [71] (2007)	Stereo visual odometry	Stereo cameras/stereo scale	No learned component; operational visual odometry	Operational mobility (Mars rover flight operations)
Yutu visual localization [16] (2014)	Cross-site visual localization	Stereo imagery/stereo geometry	No learned component; image-based localization	Operational mobility (ground-based localization in Chang’e-3 Yutu mission operations)
Lunar surface-mobility prototypes and studies.
Crater-landmark localization [50] (2022)	Particle-filter/map localization	Vision, LiDAR, crater map/map landmarks	Semantic crater detection and association	Simulation + mission data (Chang’e stereo imagery)
LunarNav [51] (2023)	Particle-filter/map localization	Vision, LiDAR, crater map/map landmarks	Semantic crater localization	Simulation
ShadowNav [49] (2023)	Particle-filter crater localization	Stereo + crater map/stereo and landmarks	Enhancement, active light, crater-edge matching	Simulation and planetary analog
LuVo [54] (2025)	Stereo visual odometry	Stereo cameras/stereo scale	LightGlue + homography matching; fail-safe logic	Simulation
DEM-anchored SLAM [52] (2026)	Visual SLAM with DEM anchoring	Stereo + DEM/stereo and map	Learned features; DEM anchoring	Simulation and planetary analog (preprint)
BEVIO [53] (2026)	Sparse-update stereo VIO (EKF)	Stereo + IMU/stereo–inertial scale	Bird’s-eye-view association; update scheduling	Simulation and planetary analog (rover analog; preprint)
Lo-SLAM [55] (2025)	Object-oriented SLAM	Stereo/object and stereo constraints	Object identification; multilevel mapping	Simulation (synthetic lunar SePT sequences)
End-to-end learned VIO methods.
VINet/causal transformer [94,95] (2017/2024)	End-to-end learned pose and fusion	Monocular + IMU/learned representation	Learned fusion and pose	Terrestrial test (method papers); no lunar evidence identified

Table 3. Representative methods, demonstrations, and evidence relevant to lunar VIO. Evidence labels state the reported validation boundary and do not imply equal maturity or flight readiness.

Representative Work (Year)	Role/Estimator	Inputs/Metric Source	Learned Component or Distinguishing Mechanism	Reported Evidence Boundary
Terrestrial reference methods (not lunar or planetary systems).
MSCKF [8] (2007)	Filter-based VIO (MSCKF)	Monocular + IMU/inertial scale	No learned component; multi-state feature constraints	Method paper; terrestrial tests; no lunar evidence
OKVIS [168] (2015)	Keyframe VIO; nonlinear optimization	Stereo + IMU/stereo–inertial scale	No learned component; tightly coupled optimization	Method paper; terrestrial tests; no lunar evidence
VINS-Mono/VINS-Fusion [11,161] (2018/2019)	Optimization-based VIO/multi-sensor odometry	Monocular or stereo + IMU/inertial or stereo scale	No learned component; multi-sensor optimization	Method/framework papers; terrestrial tests
OpenVINS [9] (2020)	Research platform; MSCKF family	Monocular or stereo + IMU/inertial or stereo scale	No learned component; configurable estimator platform	Platform paper; terrestrial tests; no lunar evidence
ORB-SLAM3 [12] (2021)	Visual–inertial SLAM; full optimization	Monocular or stereo + IMU/inertial or stereo scale	No learned component; multi-map SLAM	Method paper; terrestrial tests; no lunar evidence
DM-VIO [145] (2022)	Direct VIO; nonlinear optimization	Monocular + IMU/inertial scale	No learned component; delayed marginalization	Method paper; terrestrial tests; no lunar evidence
PIVO [102] (2018)	Probabilistic inertial–visual odometry (EKF)	Monocular + IMU/inertial scale	No learned component; probabilistic occlusion handling	Terrestrial test (WACV method paper); no planetary or flight evidence
Operational planetary localization evidence.
MER visual odometry [71] (2007)	Stereo visual odometry	Stereo cameras/stereo scale	No learned component; operational visual odometry	Operational mobility (Mars rover flight operations)
Yutu visual localization [16] (2014)	Cross-site visual localization	Stereo imagery/stereo geometry	No learned component; image-based localization	Operational mobility (ground-based localization in Chang’e-3 Yutu mission operations)
Lunar surface-mobility prototypes and studies.
Crater-landmark localization [50] (2022)	Particle-filter/map localization	Vision, LiDAR, crater map/map landmarks	Semantic crater detection and association	Simulation + mission data (Chang’e stereo imagery)
LunarNav [51] (2023)	Particle-filter/map localization	Vision, LiDAR, crater map/map landmarks	Semantic crater localization	Simulation
ShadowNav [49] (2023)	Particle-filter crater localization	Stereo + crater map/stereo and landmarks	Enhancement, active light, crater-edge matching	Simulation and planetary analog
LuVo [54] (2025)	Stereo visual odometry	Stereo cameras/stereo scale	LightGlue + homography matching; fail-safe logic	Simulation
DEM-anchored SLAM [52] (2026)	Visual SLAM with DEM anchoring	Stereo + DEM/stereo and map	Learned features; DEM anchoring	Simulation and planetary analog (preprint)
BEVIO [53] (2026)	Sparse-update stereo VIO (EKF)	Stereo + IMU/stereo–inertial scale	Bird’s-eye-view association; update scheduling	Simulation and planetary analog (rover analog; preprint)
Lo-SLAM [55] (2025)	Object-oriented SLAM	Stereo/object and stereo constraints	Object identification; multilevel mapping	Simulation (synthetic lunar SePT sequences)
End-to-end learned VIO methods.
VINet/causal transformer [94,95] (2017/2024)	End-to-end learned pose and fusion	Monocular + IMU/learned representation	Learned fusion and pose	Terrestrial test (method papers); no lunar evidence identified

Decision rules for Table 3: Evidence categories are claim-specific and non-ordinal. Method or modeling denotes an analytical formulation or algorithm without representative system validation; terrestrial test denotes evaluation using real non-planetary data or platforms; simulation denotes evaluation in a model-generated environment; planetary analog denotes a physical Earth or laboratory experiment designed to reproduce selected planetary conditions; mission data denotes analysis of actual lunar or planetary observations without a complete synchronized camera–IMU ground-truth trajectory; and operational mobility denotes use during surface mission operations. Each entry is assigned the most specific boundary directly demonstrated by the cited source; relevance alone does not justify promotion to a stronger evidence category. The categories are non-additive and are not converted into a numerical quality or maturity score. Publication status (e.g., preprint) is reported separately and does not change the evidence category.

Four cross-cutting patterns clarify what Table 3 does—and does not—show. They separate evidence modality, lunar-specific measurement design, integrated robustness, and end-to-end learning so that differences in evidence boundary are not mistaken for differences in algorithmic merit.

Pattern 1—Evidence modality. The reported validation domain limits the conclusion that can be drawn. Yutu publications provide lunar imagery and operational localization evidence [16,72,73,74,75], and crater-landmark work processes Chang’e stereo imagery [50], whereas most newer methods use simulation or Earth analogs. SPICE-HL3, BASEPROD, MADMAX, Katwijk, and Devon Island strengthen analog benchmarking [63,64,65,66,67], but no public end-to-end lunar flight camera–IMU trajectory benchmark with independent ground truth has been identified. Authentic lunar imagery can validate selected perception functions without validating the complete inertial, thermal, timing, and closed-loop chain.

Pattern 2—Lunar-specific measurement design. Lunar prototypes commonly retain established filtering, smoothing, visual odometry, or particle filter back ends while focusing on crater or DEM anchoring, geometry-aware matching, and illumination handling. This pattern differs from Pattern 1: it concerns where the technical contribution is placed, not where it was tested. The comparison indicates that dependable metric cues under lunar conditions are as important as estimator architecture; it does not imply that back-end estimation is solved.

Pattern 3—Integrated robustness. Most entries provide an individual robustness mechanism rather than a complete architecture that monitors state quality and coordinates rejection, fallback, and recovery. DM-VIO contributes delayed marginalization; ShadowNav addresses illumination; LuVo reports fail-safe logic; and BEVIO uses sparse visual updates as a resource lever. Unlike Pattern 2, this pattern concerns system-level health and recovery across components, for which a commonly validated lunar VIO design remains absent.

Pattern 4—End-to-end learning. End-to-end methods jointly learn motion and fusion representations [94,95], unlike the modular lunar prototypes in Patterns 2 and 3. Table 3 contains no lunar evaluation or explicit estimator-level health mechanism for this class, so current evidence cannot support a primary or supplementary flight role. A fair comparison requires frozen training data, explicit uncertainty, and ablation against a geometric back end using the same sensor stream; average pose error alone does not establish preserved metric scale or predictable out-of-distribution failure. The failure mode of concern for this class is not only larger error under domain shift but absent or miscalibrated confidence. The representative end-to-end methods [94,95] output relative-pose point estimates without an accompanying uncertainty, so out-of-distribution degradation is not signaled to the estimator. Where learned confidence or covariance is added, its calibration is typically established on data drawn from the training distribution; modern networks can be miscalibrated even in distribution, and post hoc recalibration on held-out data does not by itself guarantee coverage under domain shift [107,108]. The learned fusion path also offers no explicit residual interface at which such degradation can be detected or gated. Published evaluations of end-to-end VIO under controlled distribution shift remain scarce, and none under lunar conditions was identified. Until uncertainty calibration is evaluated under lunar-representative conditions, end-to-end outputs should be treated as experimental comparisons rather than as state estimates with trustworthy covariance.

6.2. Lunar Datasets and AI Sim-to-Real Validation

Sim-to-real validation requires tiered evidence. Real lunar orbital or surface products test authentic radiometry, morphology, and mission hardware in the variables they expose; Earth or laboratory analogs provide controllable sensor motion and ground truth; synthetic resources provide synchronized multi-sensor data and exact labels. These levels support different claims and should not be pooled as if they sampled the same distribution. The tiers differ in the claims they can support; they do not form an ordinal quality scale. Lunar Reconnaissance Orbiter Camera (LROC) products in the Planetary Data System (PDS) provide extensive real lunar orbital imagery, calibration levels, geometry metadata, mosaics, and derived terrain products [189]. They are valuable for illumination studies, crater or terrain detection, and orbital priors, but do not reproduce rover viewpoint, surface motion, onboard IMU measurements, or traverse ground truth. The China Lunar and Planetary Data System (CLPDS) manages Chang’e scientific, engineering, telemetry, and auxiliary product classes, with public availability depending on mission, product, and release policy [190]. Published Chang’e-3/4 mapping, localization, and teleoperation studies provide genuine lunar operational evidence [16,50,72,73,74,75], but the reviewed public products do not form a standardized rover camera–IMU trajectory benchmark with independent ground truth.

Analog and synthetic resources fill different portions of this gap. S3LI supplies a terrestrial planetary-analog multi-sensor sequence [59]. The National Aeronautics and Space Administration (NASA) POLAR Traverse Dataset provides controlled stereo exposures, calibration, camera positions, and dense geometry under simulated polar lighting [191], but no IMU stream or rover dynamics. Publicly described VIPER-related resources include the RSIM simulator and associated test infrastructure [192]; the reviewed source documents cover sensor, terrain, illumination, and slip modeling, but do not establish a downloadable, synchronized VIPER test-telemetry benchmark. It should therefore not be labeled as VIPER flight or rover telemetry. OAISYS, LunarSim, and LuSNAR [60,61,62] provide controllable synthetic evidence; LunarSim and LuSNAR include camera–IMU and ground-truth outputs, but their realism remains conditional on rendering, sensor, terrain, and dynamics models. Recent public analog resources broaden this evidence: SPICE-HL3 supplies 88 calibrated high-latitude stereo-inertial sequences with wheel odometry and precise ground truth [63]; BASEPROD combines navigation, terrain, and scientific sensing [64]; and MADMAX, Katwijk, and Devon Island provide field-scale rover navigation data [65,66,67]. Large facilities such as LUNA and associated regolith simulants support controlled integration testing [69,70]. PennCOSYVIO, ADVIO, UZH-FPV, TartanAir, Blackbird, Newer College, and the comparative VIO benchmark provide terrestrial or synthetic stress tests for calibration, transitions, aggressive motion, dynamics, and cross-modal estimation [193,194,195,196,197,198,199].

Table 4 is a claim-to-evidence guide, not an exhaustive catalog. Real lunar archives, analog datasets, synthetic resources, and terrestrial stress tests support different claims. It therefore lists only each resource’s primary VIO use and the principal limit on lunar transfer.

Table 4. Representative resources for lunar VIO development and validation. Contents and availability reflect sources accessed on 1 August 2026.

Resource	Evidence Type	Primary VIO Use	Critical Limitation
LROC PDS [189]	Real lunar; orbital	Polar imagery, calibration, geometry, and terrain products	Orbital view; no synchronized rover camera–IMU ground truth
Chang’e/CLPDS [16,50,190]	Real lunar mission	Camera/science products, telemetry classes, and operational evidence	Variable release and synchronization; no common benchmark trajectory
S3LI [59]	Earth analog	Stereo–LiDAR–IMU sequence for planetary SLAM	Earth lighting, materials, gravity, and thermal regime
POLAR Traverse [191]	Laboratory analog	Stereo exposure sweep with calibration and dense geometry	No IMU or rover dynamics; bounded terrain and lighting
VIPER RSIM/test infrastructure [192]	Synthetic; simulation and test infrastructure	Sensor, terrain, illumination, and slip modeling	No public synchronized test-telemetry benchmark
OAISYS/LunarSim/LuSNAR [60,61,62]	Synthetic lunar	Multi-sensor streams with pose, depth, or semantic labels	Model-dependent realism; no flight environmental effects
SPICE-HL3 [63]	Laboratory lunar analog	88 calibrated stereo–inertial sequences, wheel data, and ground truth	Controlled terrain/lighting; no flight dust, radiation, or thermal cycling
BASEPROD/MADMAX/Katwijk/Devon [64,65,66,67]	Field analogs	Rover sensing, wheel/IMU data, mapping, and trajectory references	Heterogeneous sensors; lunar relevance is component-specific
PennCOSYVIO/ADVIO/Newer College [193,194,198]	Terrestrial VIO	Calibration, transition, and multimodal regression tests	Non-lunar scenes, motion, gravity, and environment
UZH-FPV/Blackbird [195,197]	Terrestrial; aggressive flight	Rapid-motion and high-rate camera–IMU stress tests	Aerial dynamics differ from slow lunar rover traverses
TartanAir [196]	Synthetic terrestrial	Scalable imagery, depth, optical flow, and pose labels	Non-lunar rendering, scenes, and dynamics
MVSEC/DSEC [29,30]	Terrestrial; event camera	High-dynamic-range event/frame/inertial benchmarking	Terrestrial scenes; no lunar sensor effects
SePT/Lo-SLAM [55]	Synthetic lunar	Object labels and map-relative evaluation	No synchronized camera–IMU flight telemetry

Table 4. Representative resources for lunar VIO development and validation. Contents and availability reflect sources accessed on 1 August 2026.

Resource	Evidence Type	Primary VIO Use	Critical Limitation
LROC PDS [189]	Real lunar; orbital	Polar imagery, calibration, geometry, and terrain products	Orbital view; no synchronized rover camera–IMU ground truth
Chang’e/CLPDS [16,50,190]	Real lunar mission	Camera/science products, telemetry classes, and operational evidence	Variable release and synchronization; no common benchmark trajectory
S3LI [59]	Earth analog	Stereo–LiDAR–IMU sequence for planetary SLAM	Earth lighting, materials, gravity, and thermal regime
POLAR Traverse [191]	Laboratory analog	Stereo exposure sweep with calibration and dense geometry	No IMU or rover dynamics; bounded terrain and lighting
VIPER RSIM/test infrastructure [192]	Synthetic; simulation and test infrastructure	Sensor, terrain, illumination, and slip modeling	No public synchronized test-telemetry benchmark
OAISYS/LunarSim/LuSNAR [60,61,62]	Synthetic lunar	Multi-sensor streams with pose, depth, or semantic labels	Model-dependent realism; no flight environmental effects
SPICE-HL3 [63]	Laboratory lunar analog	88 calibrated stereo–inertial sequences, wheel data, and ground truth	Controlled terrain/lighting; no flight dust, radiation, or thermal cycling
BASEPROD/MADMAX/Katwijk/Devon [64,65,66,67]	Field analogs	Rover sensing, wheel/IMU data, mapping, and trajectory references	Heterogeneous sensors; lunar relevance is component-specific
PennCOSYVIO/ADVIO/Newer College [193,194,198]	Terrestrial VIO	Calibration, transition, and multimodal regression tests	Non-lunar scenes, motion, gravity, and environment
UZH-FPV/Blackbird [195,197]	Terrestrial; aggressive flight	Rapid-motion and high-rate camera–IMU stress tests	Aerial dynamics differ from slow lunar rover traverses
TartanAir [196]	Synthetic terrestrial	Scalable imagery, depth, optical flow, and pose labels	Non-lunar rendering, scenes, and dynamics
MVSEC/DSEC [29,30]	Terrestrial; event camera	High-dynamic-range event/frame/inertial benchmarking	Terrestrial scenes; no lunar sensor effects
SePT/Lo-SLAM [55]	Synthetic lunar	Object labels and map-relative evaluation	No synchronized camera–IMU flight telemetry

Decision rules for Table 4: A resource is classified by its acquisition setting and the variables it directly exposes, not by visual resemblance alone. Real lunar orbital or surface products may support claims about authentic radiometry, morphology, mission imagery, or the specific telemetry provided, but not trajectory-level VIO accuracy without synchronized camera–IMU data and independent ground truth. Planetary analogs may support sensor-integration and trajectory claims under the tested Earth or laboratory conditions, whereas synthetic resources support controlled pose, depth, semantic, and failure-injection studies subject to their rendering and dynamics assumptions. Terrestrial VIO or event-camera datasets support algorithm regression and stress testing but do not establish lunar transfer. The evidence types are non-ordinal and non-additive and are not converted into a common numerical score.

The representativeness gap is therefore joint rather than merely numerical: real lunar archives lack the complete synchronized state-estimation stream, whereas the most complete camera–IMU datasets are synthetic or analog. A useful benchmark suite should report raw and processed image products, hardware timestamps, camera–IMU calibration, inertial noise and bias models, trajectory ground truth independent of the tested estimator, terrain and illumination strata, motion excitation, scale handover events, and target-hardware timing. Component-only resources should support component-level claims; end-to-end VIO claims require synchronized trajectories and metric ground truth. Existing terrestrial VIO benchmarks remain useful for controlled algorithm regression, but they cannot replace lunar illumination, motion, thermal, and terrain strata [193,194,195,196,197,198,199].

A defensible sim-to-real protocol would tune on synthetic data, freeze the evaluated configuration, test it on analog sequences with independent ground truth, and then assess only the observable components on real lunar products. Performance should be reported separately at each level, including calibration and failure cases, rather than presenting the simulation-to-lunar transfer as established solely by visual similarity. Figure 8 organizes this protocol as an evidence ladder, linking each level to the metrics that can be supported.

Figure 8. Evidence ladder for lunar VIO development and evaluation. The rungs indicate domain relevance for claims about lunar operation, not methodological quality; each rung supports different claims. Synthetic scenes support controlled parameter sweeps and dense ground-truth labels; planetary analogs provide real sensors, motion, compute, and terrain; flight-derived lunar products offer the highest domain relevance but often lack synchronized inertial data or independent trajectory ground truth. Estimator configurations should be frozen prior to cross-domain transfer, and trajectory, integrity, failure-mode, and resource metrics should be reported separately at each evidence level.

6.3. Lunar AI-VIO Evaluation and System Integration

AI components occupy distinct interfaces in the estimation chain and should not be treated as interchangeable. In the representative methods below, learned models generate visual correspondences, semantic associations, depth priors, or relative-motion estimates; the downstream state estimator remains geometric or probabilistic in several hybrid systems. BEVIO is included as a model-based comparator because its bird’s-eye-view association is geometric rather than learned.

Table 5 and Table 6 separate algorithmic evidence from implementation reporting. Table 5 retains each source’s native dataset, metric, and evaluation protocol; Table 6 reports only explicitly stated hardware and computational measurements. NR denotes information not reported by the original source. Values obtained under different datasets or protocols are not ranked or treated as directly comparable.

Table 5. Source-reported comparison of representative AI components and a model-based comparator relevant to lunar VIO. Results retain the original metrics and protocols; they are not directly comparable across rows. NR = not reported. Metric abbreviations: mAP = mean average precision; AUC = area under the cumulative pose-error curve up to the stated threshold; ATE = absolute trajectory error; ARE = absolute rotation error; RMSE = root-mean-square error; RPE = relative pose error; AbsRel = absolute relative depth error; δ1 = fraction of pixels for which max(d/d*, d*/d) < 1.25, where d and d* are the predicted and ground-truth depths; IoU = intersection over union. Other abbreviations: CNN = convolutional neural network; ICP = iterative closest point; LSTM = long short-term memory; MAV = micro aerial vehicle; SAM = Segment Anything Model.

Method	AI Component and Estimator Role	Evaluation Data/Domain	Reported Metrics	Representative Source-Reported Result	Evidence Boundary/Principal Limitation
SuperPoint [19]	Self-supervised CNN detector and 256-D descriptor; generates visual measurements, not a state estimate.	Synthetic Shapes, MS-COCO and HPatches; terrestrial/planar imagery.	Repeatability, homography correctness, nearest-neighbor mAP and matching score.	HPatches: homography correctness 0.684 at 3 px; nearest-neighbor mAP 0.821.	No lunar, IMU-fusion or trajectory evaluation; failure under extreme unseen in-plane rotation was reported.
XFeat [20]	Lightweight CNN keypoint heatmap, 64-D descriptor and reliability map; visual front-end measurement generation.	MegaDepth-1500, ScanNet-1500, HPatches and Aachen; terrestrial imagery.	Relative-pose AUC, mean homography accuracy (MHA) and localization accuracy.	MegaDepth AUC@20° = 67.7; HPatches viewpoint MHA@5 px = 81.1.	No lunar or VIO trajectory evaluation; limited robustness to aggressive viewpoint changes and ambiguous image pairs.
LuVo [54]	DISK/LightGlue correspondences after homography warping; Umeyama and optional ICP remain the pose estimators.	155 simulated VIPER lunar stereo odometry pairs.	ATE RMSE, ARE RMSE and success rate (within 0.3 m and 5°).	ATE RMSE 0.081 m; ARE RMSE 0.378°; average success rate 0.867.	Pairwise simulation only; assumes locally planar terrain and does not establish long-horizon or flight performance.
BEVIO [53] (model-based comparator)	No learned component: BRISK, Hamming matching, RANSAC and geometric bird’s-eye-view warping feed the xVIO EKF.	Plaster City rover analog and DARTS lunar simulation; day/night runs.	Feature inlier ratio, 10 m RPE RMSE and minimum reliable visual-update rate.	Mean minimum reliable update rate was 76.6% of the image-space baseline; Night-2 RPE RMSE was 0.213 m at 0.25 Hz.	Approximately planar-scene assumption; sparse rates were emulated by frame skipping; not a learning-based method.
Depth Anything V2 [92]	DINOv2/DPT monocular depth prediction; a potential geometric prior, not a VIO or pose estimator.	Synthetic and large-scale terrestrial training; KITTI, NYU-D, Sintel, ETH3D, DIODE and DA-2K evaluation.	AbsRel, δ1 and sparse relative-depth-pair accuracy.	ViT-L on KITTI: AbsRel 0.074 and δ1 0.946; DA-2K accuracy 97.1%.	No lunar, temporal-consistency, IMU-fusion or pose evaluation; metric transfer to lunar imagery is unverified.
VINet [94]	End-to-end CNN-LSTM fusion that regresses visual–inertial motion increments.	EuRoC indoor MAV and KITTI road-driving sequences.	KITTI translation/orientation errors and a calibration-robustness score.	Exact tabulated trajectory result: NR; KITTI trajectory errors are presented graphically, and the calibration-table score is not sufficiently defined for cross-method comparison.	No lunar evaluation, loop closure or mapping; embedded performance was not reported.
Causal transformer/VIFT [95]	Causal transformer fuses learned visual and inertial features and estimates relative pose.	KITTI road-driving sequences 05, 07 and 10 for testing.	Relative translation drift (%) and relative rotation drift (degrees), over 100–800 m subsequences.	Sequence 07: 1.75% and 0.47°; sequence 10: 2.11% and 0.39°.	Road-motion evaluation only; lunar illumination, low texture and onboard inference were not evaluated.
Lo-SLAM [55]	Point-prompted SAM masks support target data association; geometric object-SLAM optimization estimates camera/object poses.	Four SePT simulated lunar target-approach sequences.	Mask IoU, centroid/scale error and object-centric trajectory error.	Final report: average target pose errors of 0.36 m in centroid and 0.27 m in scale, and approximately 0.49% object-centric trajectory error.	Simulation only; manual region-of-interest selection remains; detailed efficiency reporting is unavailable.

Table 6. Source-reported hardware and computational information for the methods in Table 5. Sensor acquisition or deliberately subsampled update rates are not interpreted as processing throughput. NR = applicable but not reported; NA = not applicable to the component or estimator architecture. FPS = frames per second; VGA = Video Graphics Array resolution (640 × 480 pixels); VRAM = video random-access memory.

Method	Reported Hardware	Input/Acquisition Reporting	Runtime/Throughput	Model Size and Estimator Window or Context	Deployment Reporting Limitation
SuperPoint [19]	NVIDIA Titan X GPU; descriptor sampling also timed on an unspecified CPU.	480 × 640; up to 1000 points.	Approximately 11.15 ms forward + 1.5 ms descriptor sampling; approximately 13 ms total/70 FPS.	Model size: NR. Estimator sliding window: NA (front-end component).	Desktop-GPU timing; no embedded or lunar-platform measurement.
XFeat [20]	Intel i5-1135G7 2.40 GHz CPU; Cortex-A53 test; training on RTX 4090.	VGA laptop test; 480 × 360 on Cortex-A53.	Sparse: 27.1 ± 0.33 FPS; semi-dense: 19.2 ± 1.12 FPS; Cortex-A53: 1.8 FPS.	Model size: NR; training used 6.5 GB VRAM. Estimator sliding window: NA (front-end component).	Training VRAM is not inference memory; no radiation/thermal or flight-hardware test.
LuVo [54]	Intel i9-12900HK 3.2 GHz CPU and NVIDIA RTX A2000 GPU.	Simulated VIPER stereo odometry pairs.	0.61 s per odometry pair; LuVo + ICP 2.17 s per pair.	Model size: NR. Estimator sliding window: NA (pairwise pose estimation).	Workstation result; no latency distribution or flight-target measurement.
BEVIO [53]	Onboard-computer model NR; stereo cameras and VectorNav VN100 IMU.	1024 × 1024 images at 2 Hz; IMU at 400 Hz.	Processing latency/FPS: NR. Tested 0.25–2 Hz visual-update rates were generated by frame skipping.	Model size: NA (no learned component). Estimator sliding window: NA (EKF-based state estimation).	Update frequency is not measured computational throughput.
Depth Anything V2 [92]	NVIDIA V100 GPU; exact variant NR.	Latency input resolution NR.	Small: 60 ms; Large: 213 ms.	Small/Large: 25 M/335 M parameters. Estimator sliding window: NA (single-image depth component).	No embedded or lunar-platform measurement; latency conditions are incompletely specified.
VINet [94]	Single NVIDIA Tesla K80 GPU.	Input resolution NR; EuRoC/KITTI sensor streams.	Visual CNN forward pass: 160 ms; LSTM updates > 200 Hz.	Model size: NR. Estimator sliding window: NA (recurrent sequence model, not a fixed-lag estimator).	Desktop-GPU timing; full-model size and embedded performance NR.
Causal transformer/VIFT [95]	NVIDIA GeForce RTX 4060 Laptop GPU.	KITTI camera 10 Hz and IMU 100 Hz (sensor rates).	Inference latency/FPS: NR.	Parameters: NR. Transformer context/window used for deployment: NR.	Sensor rates must not be interpreted as estimator throughput.
Lo-SLAM [55]	RTX 4090 reported for the open precursor; final-article CPU/RAM NR.	Four simulated SePT sequences.	Runtime/FPS/latency: NR.	Model size and optimization-window size: NR.	Available hardware detail comes from the precursor; final deployment cost remains unreported.

Table 6 exposes a reporting gap rather than a complete resource ranking. None of the eight methods in Table 5 and Table 6 reports processor, inference latency or throughput, peak inference memory, measured power, estimator window or context, and numerical precision together. In particular, none of the eight sources reports measured power consumption, peak inference memory, or numerical precision. A sliding-window size is not applicable to stand-alone feature or depth components, pairwise pose estimation, or EKF-based state estimation; where an architecture has a relevant context or optimization horizon but no value is provided, the entry is marked NR. Workstation timings and sensor acquisition or subsampled update rates do not establish onboard flight readiness. Consequently, no normalized SWaP score or ranking is assigned.

Read together, Table 3, Table 4, Table 5 and Table 6 show two independent requirements. Candidate system components must be compatible at the estimator level, and claims about those components must be supported by an evaluation resource appropriate to the claimed operating condition. Mature filtering or smoothing back ends provide useful baselines, while learned features, semantic constraints, illumination handling, and multiple metric cues remain configuration-dependent options. The evidence does not identify a single flight-grade default; instead, it supports evaluation across explicit terrain, illumination, motion, sensor, and SWaP conditions. System-level comparisons should consequently be organized around matched claims: front-end studies need measurement-level metrics; estimator studies need trajectory and consistency metrics; robustness studies need fault-response metrics; and deployment studies need target-hardware measurements. A strong result at one level cannot substitute for evidence at another. The most informative experiment would hold the back-end and sensor inputs constant while ablating each learned or auxiliary component across synthetic, analog, and real-product strata. This design would reveal whether observed gains arise from the learned module, additional metric information, or differences in tuning and evaluation conditions.

Table 7 organizes these requirements into a recommended evaluation matrix for lunar AI-VIO studies; a strong result at one level does not substitute for evidence at another.

Table 7. Recommended evaluation matrix for lunar AI-VIO studies, organized by evaluation level. The matrix consolidates the metrics discussed in Section 3, Section 4, Section 5, Section 6 and Section 7; reporting examples refer to the reviewed sources in Table 5 and Table 6 where such metrics are reported. Metric abbreviations follow Table 5; NIS = normalized innovation squared; NEES = normalized estimation error squared.

Evaluation Level	Recommended Metrics	Claim the Metrics Can Support	Reporting Examples from the Reviewed Sources
Front end (measurement quality)	Feature repeatability; matching score or nearest-neighbor mAP; homography or relative-pose accuracy (MHA, AUC); inlier ratio; depth error (AbsRel, δ1)	Availability and reliability of visual measurements under lunar photometry, texture, and geometry	SuperPoint, XFeat, and Depth Anything V2; feature inlier ratio in BEVIO (Table 5)
Estimator (trajectory and consistency)	ATE/RPE RMSE; scale error; normalized innovation or estimation error (NIS/NEES); uncertainty coverage; conditioning indicators such as the minimum eigenvalue of $\Lambda$ (Section 3.1 and Section 5.2) and the excitation–bias ratio (Equation (6))	Metric accuracy, calibrated uncertainty, and metric-scale integrity (Section 4.1)	LuVo (ATE/ARE RMSE, success rate), BEVIO (10 m RPE), VIFT (relative drift) (Table 5)
Robustness and failure handling	Detection delay; false-alarm rate; bounded state growth; transition continuity; recovery behavior under induced faults	Fault response and failure recovery within the estimator or at its measurement interfaces (Section 5)	None of the representative methods in Table 5 reports these metrics; see the FDIR survey [116] and the mechanisms reviewed in Section 5
Deployment (target hardware)	Processor; measured latency or throughput; memory; power; estimator window or context; numerical precision	Onboard feasibility under SWaP limits	Table 6 (with the reporting gap noted there)

7. Open Challenges and Research Directions for AI-Enhanced Lunar VIO

The preceding technical sections identify recurring limitations in evidence, metric compatibility, robustness, and deployment assessment. This section consolidates the cross-cutting challenges and research directions required to establish lunar-domain validity, calibrated uncertainty, estimator compatibility, recoverable failure behavior, and SWaP-bounded implementation. The emphasis is on testable evidence and technical dependencies, so prospective sensing and learning methods are evaluated against measurable uncertainty, interfaces, failure behavior, and hardware constraints.

7.1. Cross-Cutting Challenges

Read together, the front-end, estimator, robustness, and deployment gaps resolve into three cross-cutting challenges.

Validation Evidence and Benchmarking: The central validation problem is not a simple shortage of data but a mismatch between what each evidence source can establish. Established terrestrial VIO benchmarks support reproducible trajectory comparisons under instrumented conditions [56,57,58], planetary and lunar simulation resources expose selected terrain and sensing factors [59,60,61,62], analog traverses provide coupled vehicle–environment effects [64,65,66,67,68,69,191], and orbital or mission archives preserve genuine lunar appearance and topography [189,190]. None alone demonstrates end-to-end surface navigation across lunar illumination, dust, thermal variation, vibration, timing error, and long-duration drift. This distinction matters because an aggregate accuracy figure can hide the conditions that trigger estimator inconsistency, relocalization failure, scale loss, or unacceptable compute load. Evidence should therefore remain labeled by modality and by the state variables directly observed; image realism, for example, does not establish inertial fidelity or closed-loop trajectory accuracy.

A credible benchmark should specify sensor models and synchronization, intrinsic and extrinsic calibration, exposure and timestamp behavior, inertial noise and bias, terrain and illumination strata, motion excitation, fault injection, independent ground-truth uncertainty, and target-hardware runtime, memory, energy, and latency. Public development data should be separated from frozen calibration and hidden test sequences, with predeclared metrics for absolute and relative trajectory error, scale drift, uncertainty coverage, tracking availability, failure-detection delay, recovery continuity, and resource use. Results should include per-stratum distributions and failure counts rather than only mean performance, and repeated tuning on the test domain should be disclosed. Progress would be evidenced by independently repeatable rankings across versioned splits, bounded degradation across synthetic, analog, and lunar-product tests, and conclusions that remain stable when algorithm settings and compute budgets are fixed—not by a single improved score in a single domain.

Lunar-Domain Transfer: Lunar-domain transfer is consequential because learned features, depth, segmentation, enhancement, and confidence models can preserve terrestrial benchmark performance while altering their error structure under unfamiliar reflectance, illumination, terrain texture, or sensor response conditions. Lunar photometry and polar thermal conditions are physically distinctive [13,14,15], whereas transfer learning and cross-dataset depth studies show that generalization depends on the source and target distributions rather than automatically following from model scale [89,90,91,167]. The reviewed lunar low-light and permanently shadowed-region studies provide task-specific perception evidence [27,28], but they do not by themselves establish calibrated VIO measurements, metric scale, or estimator consistency. Consequently, apparent qualitative recovery, feature count, or segmentation accuracy should not be treated as navigation benefit unless the resulting residuals and state estimates are evaluated.

Transfer studies should report training and pretraining sources, synthetic-to-real proportions, target-domain samples used for adaptation, frozen hyperparameters, camera response and radiometric processing, and results stratified by incidence angle, dynamic range, texture, terrain type, motion, and sensor temperature. At the estimator interface, learned outputs require geometric validity checks, covariance or confidence calibration, out-of-distribution indicators, and ablations against non-learned measurements under the same compute budget; calibration error and uncertainty coverage should accompany task accuracy [107,108]. Relevant failure modes include confident mismatches, depth-scale bias, hallucinated structure after enhancement, and adaptation that improves one stratum while degrading another. Progress would mean lower transfer degradation on held-out lunar-like conditions, calibrated confidence that predicts navigation error, and reproducible estimator-level gains after configuration freezing. The evidence boundary should remain explicit: success on synthetic or orbital imagery motivates further testing but does not establish surface-operational robustness.

Flight Readiness and SWaP-Constrained Integration: The gap to flight readiness arises because component accuracy, integrated estimator behavior, and implementability are different claims. Embedded vision studies and space-edge reviews show that processor architecture, memory movement, numeric precision, and radiation-tolerant implementation constrain achievable performance [17,18,150]. Quantization and distillation can reduce neural inference cost [147,148], but compression may also alter feature repeatability, confidence calibration, or rare-condition behavior. Likewise, a front end that meets the average frame rate can still violate worst-case latency, memory, or energy limits when image content, event rate, optimization iterations, or recovery processing increases. These interactions matter in VIO because delayed or dropped measurements, thermal throttling, and unsynchronized duty cycling can alter estimator observability and uncertainty, rather than merely slowing the pipeline.

Reporting should therefore couple trajectory and consistency metrics with processor type, precision, clocking, peak and sustained memory, per-stage latency distributions, energy per update, sensor duty cycle, start-up and reinitialization cost, and worst-case load. Tests should include long runs, timing jitter, dropped frames, compute saturation, sensor faults, and transitions between nominal and degraded modes; fault-detection and recovery claims should be aligned with established space FDIR concerns [116]. Comparisons need a common hardware and accuracy envelope so that lower power is not credited when measurement quality or update rate has changed without disclosure. Progress would consist of an integrated, frozen pipeline meeting declared worst-case deadlines and resource margins on representative hardware while retaining calibrated uncertainty and recoverable failure behavior across the tested strata. Such evidence supports increasing implementation maturity, but it should not be described as flight readiness without environmental qualification, mission-specific assurance, and representative system-level validation.

7.2. AI-Enhanced Lunar VIO Research Directions

Four prospective directions connect multiple parts of the VIO pipeline. They are framed as testable research priorities supported by enabling evidence, not as demonstrated lunar navigation solutions.

Continuous-Time Estimation: Continuous-time trajectory representations provide a principled interface for asynchronous events, rolling-shutter observations, multi-rate inertial data, and measurements with non-coincident timestamps. Spline-based VIO and efficient Lie-group derivatives provide established formulations [200,201]. However, the interface must specify the trajectory state, knot placement, measurement support, bias evolution, clock offsets, and how covariance or information is transferred to downstream mapping, control, and health monitoring. This action is particularly important because temporal calibration and camera–IMU observability are coupled [44,154,157]: a flexible trajectory can absorb timing or calibration error rather than identify it, producing visually plausible estimates with overconfident uncertainty. Initialization and bias estimation remain separate failure points [45,46], and continuous-time modeling does not remove weak excitation or poor visual geometry. Computational and memory cost is a first-order constraint for this direction. Space-grade processors offer substantially lower throughput and memory than the commercial hardware on which the formulations above are typically evaluated [17,18], and the per-update cost of continuous-time optimization grows with knot density and measurement rate. Table 6 contains no continuous-time estimator, but its source-reported figures illustrate the gap between workstation and embedded execution: XFeat runs at 27.1 FPS on a laptop CPU at VGA resolution but at 1.8 FPS on a Cortex-A53 at 480 × 360 [20], and the latter is not a space-grade processor. A single computational and memory estimate for this research direction would therefore not be evidence-based; the evaluation directions in Section 7.1 and Section 7.2 instead call for latency, memory, and energy or power to be quantified under explicit compute limits.

Evaluation should compare continuous- and discrete-time estimators using identical sensor data, calibration priors, optimization windows, compute limits, and output rates, with ablations for knot density and adaptation policy. Required measurements include trajectory and scale error, timestamp and extrinsic estimation error, normalized innovation or estimation error, uncertainty coverage, latency, memory, solver iterations, and sensitivity to event bursts, gaps, jitter, bias change, and marginalization. Public VIO comparisons and initialization benchmarks provide useful protocol precedents [158,199], although their domains do not establish lunar performance. Data should include accurately time-stamped high-rate motion with independent trajectory truth and deliberately perturbed timing and calibration. Progress would be a repeatable accuracy–consistency or accuracy–compute improvement in the strata where asynchrony matters, without degradation elsewhere. Evidence should remain bounded to the tested sensor rates and motion regimes; modeling convenience alone is not proof of a lunar advantage.

Event and Thermal Sensing: Event and thermal sensing address different observability gaps and should not be grouped as universally low-light sensors. Event cameras provide high temporal resolution and dynamic range, but measurements require incident light and temporal contrast, and their rate depends on scene motion, threshold variation, and noise [202]. Thermal cameras can operate without visible illumination, yet usable geometric contrast depends on terrain temperature structure, optics, resolution, non-uniformity correction, and thermal drift [203]. Existing thermal–inertial, thermal-edge, hybrid frame–event–IMU, and event-only systems demonstrate workable estimator architectures in terrestrial settings [31,32,33,34,35,36,37]; event datasets support controlled comparison [29,30]. These studies establish feasibility and interfaces, not general performance in permanently shadowed regions of the Moon. Low-light enhancement and recovered PSR image information [27,28] are complementary forms of perceptual evidence but do not replace direct navigation evaluation.

Each modality should expose timestamped measurements, calibration state, valid operating range, and uncertainty to the estimator. Required datasets should pair frames, events, thermal imagery, and IMU data with independent trajectory truth; record illumination, temperature, exposure, event thresholds, motion, scene contrast, and sensor self-heating; and include static darkness, low-texture terrain, rapid shadow transitions, saturation, hot pixels, event bursts, and prolonged low-event intervals. Comparisons should include visible-only, event-only, thermal-only, and fused configurations under matched compute and power budgets, reporting calibration stability, trajectory error, uncertainty coverage, tracking availability, latency, bandwidth, and incremental SWaP. Failure handling should reject or down-weight stale, bursty, or weakly calibrated measurements without concealing loss of observability. Progress would be a statistically supported extension of the tested operating envelope with calibrated estimator uncertainty and acceptable incremental resource cost. It would not establish all-illumination navigation beyond the measured thermal and photometric conditions.

Tiered Simulation-to-Reality Validation: A tiered protocol is needed because synthetic, analog, orbital, and surface-derived evidence answer different questions. Simulation permits controlled variation and exact state truth [60,61,192], analog datasets expose physical optics, rover motion, and terrain interactions [63,64,65,66,67,68,69,191], and lunar archives provide genuine appearance or topography [189,190] but often lack synchronized surface IMU and trajectory truth. General VIO datasets and comparisons can stress motion, initialization, and algorithmic robustness [158,193,194,195,196,197,198,199], while planetary visual-odometry studies add relevant terrain evidence [59,146]. Combining these sources is valuable only if domain boundaries remain visible; pooling them into one score can mistake synthetic coverage for physical validity or image realism for estimator validation.

The protocol should freeze code, weights, calibration policy, and decision thresholds before each transfer stage; reserve adaptation data separately; and predefine pass, fail, and abstention criteria. Sensor and rendering parameters, random seeds, ground-truth uncertainty, analog-site geometry, and any post hoc exclusions should be versioned. Comparisons should use common sensor subsets and compute budgets, then report error, consistency, availability, failure, and recovery by domain and operating stratum, together with the performance change between stages. Deliberate mismatches—photometry, blur, dust-like occlusion, inertial bias, timing error, terrain frequency, and thermal conditions—are needed to identify which assumptions fail. Progress would be narrowing and explaining the transfer gap on frozen tests, reproducing it across independent sites or generators, and predicting which configurations will fail before exposure to the next tier. Passing a lower tier justifies the next experiment; it does not certify performance in a higher-evidence domain.

Evidence-Driven Research Priorities: Evidence-driven prioritization should target the uncertainties that currently prevent comparison and integration, rather than ranking technologies by novelty. The priority is measurement infrastructure: synchronized camera, IMU, and candidate auxiliary-sensor data; traceable calibration and ground truth; explicit illumination, terrain, motion, timing, and fault strata; and representative compute profiles. The next is estimator compatibility: learned or alternative-sensor outputs should be compared with model-based baselines, supplied with calibrated covariance or confidence [107,108], and tested for geometric and metric validity before end-to-end gains are interpreted. Robust estimation methods provide well-founded tools for rejecting or bounding contaminated measurements [176,177], but their thresholds, false rejections, and behavior during sustained degradation still require system-level measurement. These priorities prioritize evidence quality over unsupported claims about a particular architecture.

Within frozen experiments, front-end accuracy should be linked to residual quality and state error; estimator studies should isolate scale, timing, calibration, and initialization; and integration studies should quantify worst-case latency, memory, power, and recovery. Quantization and distillation [147,148] should be evaluated for calibration and rare-condition regressions, as well as for throughput. At the same time, fault-management claims should report detection delay, false alarms, bounded state growth, and transition continuity in a form consistent with space FDIR practice [116]. Candidate event, thermal, adaptation, or continuous-time modules should advance only when they improve a declared metric without violating the test’s uncertainty or resource limits. Progress is therefore cumulative and falsifiable: shared data and protocols, reproducible component effects, estimator-level consistency, representative-hardware margins, and clearly bounded system evidence. Where these conditions are absent, the appropriate outcome is an identified evidence gap, not an extrapolated claim of operational benefit.

8. Conclusions

This review examined AI-enhanced visual–inertial navigation for post-landing lunar surface mobility as a coupled sensing, estimation, robustness, validation, and deployment problem. The literature supports VIO as a useful GNSS-denied state-estimation framework, but it does not support a single flight-ready architecture or a universally superior front end. Lunar illumination, terrain, motion, sensor configuration, calibration, and compute constraints interact, while much of the available evidence remains synthetic, analog, orbital, or terrestrial. Conclusions must therefore remain bounded by the measurement, state variables, operating strata, and validation level actually observed.

AI is most defensible as a measured component within an explicit estimator: learned features, depth or semantic priors, illumination handling, confidence estimates, and health classifiers may improve particular measurements, but terrestrial benchmark rank is not lunar validity. Their outputs require lunar-domain transfer tests, calibrated uncertainty, geometric and metric compatibility, and ablation against model-based baselines. At the estimator level, reduced gravity alone does not make monocular scale unobservable; practical metric integrity depends jointly on motion excitation, inertial noise and bias, visual geometry, estimation horizon, and the calibration and association quality of stereo, wheel, crater, or map cues. Scale drift, source handover, and re-initialization should consequently be evaluated as coupled estimator properties rather than attributed to one sensor in isolation. Estimator-level robustness has stronger foundations in residual and innovation gating, robust costs, conditioning checks, constrained updates, and bounded-window estimation than in learned supervisory policies. Lunar evidence remains limited for calibrated learned uncertainty, integrated sensor-health fallback, adaptive resource configuration, and recovery across metric-source transitions. These mechanisms should be evaluated based on fault-detection delay, false alarms, uncertainty coverage, accumulated state error during degradation, and recovery continuity. AI-assisted decisions are credible only when tied to observable health evidence, bounded estimator responses, and an explicit degraded or invalid state.

The principal evidence gap is thus both infrastructural and algorithmic. Table 3 identifies the validation boundaries of representative systems, and Table 4 shows that current datasets and simulators support complementary claims but do not enable a common end-to-end comparison. A tiered benchmark should combine synchronized sensing, independent truth, terrain and illumination strata, excitation and handover events, fault annotations, and target-hardware timing while preserving separate synthetic, analog, and lunar-product evidence. Evidence-driven priorities should begin with benchmark and calibration foundations, advance to uncertainty-compatible estimator integration, and culminate in resource-aware flight validation and recovery. This dependency ordering keeps SWaP, robustness, and transfer claims testable and positions AI as a constrained contributor to navigation integrity rather than a substitute for estimator design or evidence.