The landscape of
GNSS-denied navigation.

Overview

GNSS is the dominant absolute reference for position and time because it is globally available and inexpensive at the receiver. The engineering problem is not accuracy in benign conditions; it is fragility. Received satellite signals are extremely weak at the antenna, the propagation environment near the ground is rich in reflections, and many systems have been built with the implicit assumption that GNSS is continuously trustworthy. That assumption leaks into everything from estimator tuning to mission logic and timing discipline.

Denial and deception are different failure classes. Jamming raises the effective noise floor and collapses tracking—often abruptly and without an informative signature beyond loss of lock. Spoofing can be worse: counterfeit signals can drive a receiver to a coherent but false solution that looks healthy until cross-checked against independent observables. Multipath and non-line-of-sight reception are not adversarial, but they create biases that can resemble slow spoofing, especially in urban and maritime environments where reflection geometry changes rapidly.

The shift from GNSS augmentation to periods of GNSS independence is therefore architectural. Robust systems use diversity of physics to bound error growth, and they treat every aid source—including GNSS—as a measurement with failure modes. That implies explicit integrity monitoring, conservative uncertainty management, and tight time synchronization across sensors. It also elevates map data and provenance to operational concerns: terrain and magnetic methods depend on prior surveys, update pipelines, and permissions—issues that matter directly in European programs pursuing sovereign autonomy without dependence on external infrastructure.

Key topics

Visual odometry and SLAM

Visual odometry (VO) and SLAM estimate motion by tracking scene structure in camera measurements. Their output is fundamentally relative: change in pose over time. Without an external reference—map correlation, known landmarks, or GNSS—they cannot provide global truth, only a locally consistent trajectory whose error accumulates with distance and time.

Monocular VO has an inherent scale ambiguity; metric scale becomes observable only through additional assumptions or inertial aiding. Visual–inertial odometry (VIO) addresses this by fusing camera and IMU, but scale and bias observability remain conditional on trajectory excitation and on accelerometer calibration. Stereo and multi-camera rigs provide direct metric depth from parallax, trading algorithmic ambiguity for mechanical stability and calibration burden; depth constraints also weaken quickly with range unless baseline is large.

Loop closure is the mechanism that can bound drift by enforcing consistency when revisiting previously seen areas. Two constraints matter operationally: many missions do not revisit, and false loop closure is a catastrophic failure mode. A single incorrect place match can inject a topological error that an optimizer will distribute through the entire map, yielding a confident but wrong solution unless robust outlier handling and integrity monitoring are engineered in.

Field performance is dominated by environmental observability. VO degrades in low-feature scenes (snow, desert, open water, bare walls), in repetitive structure (corridors, crop rows, facades), and in dynamic scenes where tracked features are not static. Lighting changes the measurement model: glare, deep shadow, flicker, and night operations can create intermittent tracking and unmodelled biases. Platform effects—vibration, rolling-shutter distortion, lens contamination, and thermal focus shift—are routine, not exceptional.

Compute and fusion architecture determine whether the method is viable as a primary aid. Front-end robustness (features, matching, outlier rejection) and back-end optimization compete for latency and power with perception and control. Tightly coupled VIO, where raw features update the inertial state, is usually more robust under partial degradation than loosely coupled fusion of black-box pose outputs, but it demands tight time alignment and stable camera–IMU calibration. Those integration costs are often underestimated in early prototypes.

Terrain-aided navigation

Terrain-aided navigation (TERCOM/TRN) bounds position error by correlating measured terrain signatures against stored elevation maps. The observable is the spatial pattern of terrain variation along track, typically measured with radar altimetry, laser ranging, or combinations of barometric altitude and terrain clearance models. An inertial solution propagates the state; correlation provides absolute updates when the terrain signature is distinctive.

Performance is governed by map quality and terrain distinctiveness. Digital elevation models must have appropriate resolution, vertical accuracy, and consistent geodetic referencing; survey artifacts and datum mismatches translate directly into biased updates. Correlation works in structured terrain—mountains, coastlines, escarpments—and collapses in flat deserts, ice plains, and open ocean where many candidate locations look alike.

TRN does not eliminate inertial drift; it corrects it intermittently. Between updates the system still drifts, and the update cadence is constrained by how quickly the observed profile becomes unique. Computational load scales with the search region: as uncertainty grows, correlation must evaluate more hypotheses, increasing latency and potentially destabilizing the estimator if updates arrive late or are overconfident. Altimeter biases and surface-dependent measurement errors are common failure drivers and must be modelled explicitly.

Bathymetric variants apply the same principle underwater using depth sounders or pressure sensors against seafloor maps. Their applicability is geography-dependent: distinctive seabed features support correlation; deep, smooth basins do not. Operationally, the method’s strengths—passive, hard to spoof at scale—are balanced by map logistics, storage, security, and licensing constraints that often dominate commercial adoption decisions.

Neuromorphic sensing

Event-based cameras report asynchronous brightness changes rather than fixed-rate frames. For navigation, they are best understood as a sensing modality that can extend the operating envelope of visual motion estimation: microsecond timestamps reduce latency, and high dynamic range helps maintain observability across harsh lighting transitions that saturate conventional cameras.

The limitations follow directly from the measurement principle. When the scene is static or low-contrast, few events are generated and motion becomes weakly constrained. Many pipelines reconstruct pseudo-frames or time surfaces to reuse conventional algorithms, reintroducing design choices about accumulation windows, noise filtering, and latency. Background activity, hot pixels, and lighting flicker can create coherent false motion if not handled carefully.

Maturity is uneven. Research results are strong in high-dynamics regimes, but toolchains, calibration procedures, and long-duration field validation are less mature than for conventional imaging. In operational systems today, event cameras are most credible as a complementary input to VIO rather than as a stand-alone navigation sensor.

Inertial navigation systems

An inertial navigation system (INS) mechanizes gyro and accelerometer measurements to propagate attitude, velocity, and position without external signals. It is the continuity backbone in GNSS-denied architectures because it is largely environment-agnostic. The cost is unbounded drift. INS error growth is not mysterious: it is the predictable integration of bias, scale, misalignment, and stochastic noise into the state.

Gyro bias drives attitude error; attitude error rotates gravity into the horizontal plane, creating a fictitious horizontal acceleration that integrates into velocity and position. Accelerometer bias integrates directly into velocity error. Because these mechanisms integrate once or twice, small improvements in bias stability can yield disproportionate improvements in position holdover over operationally relevant intervals.

Sensor grade matters because bias stability and noise density differ by orders of magnitude across MEMS, tactical, and navigation-grade technologies. Low-cost MEMS can be acceptable with frequent aiding but drift quickly when unaided and are sensitive to temperature and vibration. Tactical-grade systems extend viable outage windows at higher SWaP and cost. Navigation-grade units enable materially longer unaided performance but are expensive, physically larger, and often constrained by supply chain and exportability—an explicit consideration in European programs aiming to reduce external dependency.

Calibration and environment determine whether lab-identified models hold in the field. Temperature compensation, scale factor stability, g-sensitivity, and vibration rectification errors routinely dominate. Shock events can introduce bias steps that violate simple random-walk assumptions. Alignment is also an operational mode: autonomous systems that must cold-start or recover from resets need a defined pathway to re-establish attitude and bias estimates without human intervention.

Fusion architecture turns predictable drift into manageable drift. Loosely coupled systems fuse aiding as position/velocity updates and are modular, but they can fail abruptly when the aiding module becomes unreliable. Tightly coupled estimators update directly from raw measurements (camera features, altimeter traces, ranges), improving observability under partial degradation and enabling better residual-based integrity monitoring. The common field failure is not “INS failure” but an estimator that becomes overconfident and cannot recognize that its updates are wrong.

Magnetic navigation

Magnetic navigation uses spatial variation in the geomagnetic field or local anomalies as a map-matching reference. Like TRN, it correlates measured signatures against a stored map to provide absolute updates. Without a prior map, a magnetometer is primarily an attitude aid; positioning capability is location-dependent and map-bound.

At regional scales, anomaly fields arise from geology and can be stable enough to support correlation in areas with distinctive structure. Indoors, the dominant features are man-made: steel, rebar, machinery, and power infrastructure. Indoor magnetic fingerprinting can work in controlled environments but is brittle under structural changes, moving ferromagnetic objects, and altered electrical loads.

Interference is the operational constraint. Motors, power electronics, and payloads generate fields that must be calibrated (hard-iron/soft-iron effects) and can change with configuration. External interference—from vehicles, power lines, or deliberate field generation—can bias measurements in ways that are difficult to separate from true spatial anomalies without additional observables.

Physics also sets limits: as altitude increases, local anomalies blur and distinctiveness decreases. This makes magnetic navigation more plausible for ground and low-altitude platforms than for high-altitude flight. In practice, its most defensible role is as an opportunistic absolute update source within a fused system, not as a universal replacement for GNSS.

Accuracy claims are meaningless without stating the conditions under which error growth is bounded. Perception-based navigation can deliver low short-term error but has unbounded drift and abrupt failure modes (tracking loss, false associations) that are environment-driven. Map correlation methods can bound error where terrain or anomaly structure is distinctive, but they fail geographically and degrade when the prior map is stale or mis-referenced.

Environmental robustness differs by modality. INS is insensitive to external scene content but sensitive to temperature, vibration, and shock. VO depends on texture and lighting and is vulnerable to obscurants and lens contamination. TRN depends on sensor geometry and surface interaction (radar/laser returns) and fails in flat regions. Magnetic navigation depends on electromagnetic cleanliness and on stable anomaly structure; it is strong in some locations and unusable in others.

Compute and integration effort also separate the methods. VO/VIO pushes heavy front-end and optimization workloads onto edge compute, with tight timing constraints. TRN and magnetic correlation shift the burden into map storage, search over hypotheses, and map production logistics. Fusion complexity grows nonlinearly with each added aid source because every sensor adds calibration parameters, time alignment paths, and new failure hypotheses that must be monitored.

In credible GNSS-denied systems, no single modality carries the mission. A common pattern is an inertial core, a perception-based aid for local motion when the scene supports it, and a map-referenced aid (terrain or magnetic) to provide occasional absolute corrections—backed by integrity monitoring that can reject inconsistent updates and declare uncertainty when absolute position is not trustworthy.

Market direction

Across defense and high-consequence commercial autonomy, GNSS vulnerability is being treated less as an edge case and more as a design condition. Specifications are shifting from best-case accuracy to performance under defined outage and deception scenarios, including requirements for detection, bounded error growth, and safe degradation when absolute position cannot be trusted.

The same operating conditions exist beyond the battlefield. Offshore energy platforms in the North Sea operate in documented jamming environments from nearby vessels and land-based transmitters. Arctic and polar survey operations contend with degraded satellite geometry and ionospheric distortion that makes GNSS-only navigation unreliable for extended periods. Underground mining, deep-sea operations, and port logistics in dense urban environments all present conditions where satellite availability is intermittent or absent by geometry alone. The engineering problem — bounded position in environments where GNSS cannot be trusted — is structurally identical across these domains, even if the threat model differs.

Defense procurement is increasingly centered on assured PNT behavior: quantified uncertainty, integrity monitoring, and demonstrated resistance to both jamming and spoofing. Terrain-referenced and high-grade inertial approaches remain relevant where mission geography and budgets support them, while vision-based approaches are valued for being passive and difficult to interfere with electronically—provided they are qualified for weather, obscurants, and platform dynamics rather than curated datasets.

European sovereign autonomy adds another axis: control over sensors, timing components, and—critically—over geospatial priors and their update pipelines. The practical meaning of “infrastructure-independent autonomy” is not only operating without satellite signals, but doing so without depending on external data permissions, foreign supply chains, or opaque trust anchors that can be withdrawn when geopolitical conditions change.

Edge compute is enabling more sophisticated onboard estimation, but it does not eliminate integration discipline. Power and thermal constraints limit sustained acceleration; safety cases are harder when timing and performance depend on resource contention. Systems that field well will be those that treat GNSS as one fallible measurement among several, fuse sensors with conservative uncertainty, and verify failure modes in representative environments rather than on idealized test ranges.