The landscape of
GNSS-denied navigation.

Overview

GNSS dominates as the go-to absolute reference for position and time. It's globally available and dirt cheap at the receiver. But here's the real engineering problem: fragility. Satellite signals arrive at the antenna already faint, ground-level propagation bounces them off everything nearby, and most systems were designed with a hidden assumption that GNSS will always be there and always be trustworthy. That assumption spreads into everything: filter tuning, mission logic, synchronization discipline. Once you notice it, you see it in every subsystem.

Denial and deception are fundamentally different failure modes. Jamming pushes up the effective noise floor and tracking collapses, often suddenly and without any useful diagnostic beyond loss of lock. Spoofing is worse. Counterfeit signals can steer a receiver to a solution that's internally consistent and looks perfectly healthy until you check it against independent sensors. Multipath and non-line-of-sight reception aren't adversarial, but they introduce biases that can look like slow spoofing, especially in urban canyons and at sea where reflection geometry shifts constantly.

Moving from GNSS augmentation to periods of GNSS independence isn't a software tweak; it's an architectural requirement. Systems that hold together use multiple physics to cap error growth, and they don't pretend GNSS is special. Every aid source (GNSS included) gets treated as a measurement with specific failure modes. That means explicit integrity monitoring, conservative uncertainty budgets, and time synchronization you can count on across all sensors. It also bumps map data and source provenance to operational status: terrain methods and magnetic methods need current surveys, working update pipelines, and the legal right to use them. These aren't afterthoughts, especially in European programs building autonomous systems they control without depending on external infrastructure.

Key topics

Visual odometry and SLAM

Visual odometry and SLAM track scene structure in camera images to estimate motion. What you get is always relative: pose change from frame to frame. Without an external anchor (map correlation, known landmarks, or GNSS) you get a trajectory that's internally consistent locally but error grows without bound. Distance traveled and elapsed time both work against you.

Monocular VO can't figure out absolute scale from image alone. You need extra assumptions or inertial help. Visual-inertial odometry fuses camera and IMU to fix that, but you're not home free: scale and bias stay observable only if your trajectory moves in enough directions and the accelerometer is actually calibrated. Stereo and multi-camera systems get metric depth directly from parallax. Trade-off: you avoid the scale ambiguity but add mechanical complexity and calibration overhead. Depth also deteriorates with range unless your baseline is substantial.

Loop closure (detecting when you've returned to a place you've seen) can cap drift by enforcing consistency. Two real constraints: many missions never actually revisit, and false loop closure is catastrophic. One bad place match corrupts the whole map, and the optimizer will spread that error everywhere, leaving you with a beautifully confident but completely wrong solution. Unless you engineer in strong outlier rejection and integrity checks, it will kill you.

Field performance hinges on what the environment actually offers to observe. VO falls apart in low-texture scenes: snow, desert, open water, blank walls. Repetitive structure kills it too: corridors, crop rows, office facades all look the same. Dynamic scenes where features aren't static are a problem. Lighting matters more than you'd think. Glare, deep shadows, flicker, night operation: any of these can break tracking intermittently or introduce biases you didn't model. Platform issues like vibration, rolling-shutter artifacts, lens gunk, and thermal focus creep aren't edge cases. They're routine.

Whether this actually works as a primary aid depends on compute and fusion architecture. Front-end quality (feature detection, matching, outlier handling) competes with back-end optimization for latency and power, both fighting for resources with perception and control. Tightly coupled VIO, where raw features directly update the inertial state, usually holds up better under partial sensor loss than loosely coupled fusion of black-box pose estimates. But tight coupling demands time alignment you can trust and stable camera-IMU calibration. These costs get underestimated constantly in early designs.

Terrain-aided navigation

Terrain-aided navigation caps position error by matching measured terrain signatures to stored maps. What you measure is the spatial pattern of elevation along your path, typically with radar altimetry, laser ranging, or combinations of barometric pressure and terrain clearance estimates. An inertial filter carries the state forward. When the observed signature is distinctive enough, correlation gives you absolute corrections.

Success comes down to map quality and how distinctive the terrain is. Your digital elevation model needs the right resolution, vertical accuracy, and consistent geodetic reference. Survey flaws and datum shifts become biases in your fixes. Correlation works beautifully in structured terrain: mountains, coastlines, escarpments. Flat deserts, ice sheets, open ocean? Many locations look identical. The method collapses.

TRN doesn't kill inertial drift; it corrects it in bursts. Between updates you're still drifting, and how often you get updates depends on when the observed terrain profile becomes unique. Compute load scales with your uncertainty: bigger search region means more hypotheses to evaluate, more latency, and risk of destabilizing the filter if updates arrive late or overconfident. Altimeter biases and surface-specific measurement errors are common failure modes and need explicit modeling.

Underwater, the same idea works with depth sounders or pressure sensors matched against seafloor maps. Geography determines whether it works at all: distinctive seabed features support correlation; smooth, deep basins don’t. The method has real operational strengths: it’s passive and hard to spoof at scale. But map logistics, storage, security, and licensing issues often kill adoption before you get to field trials.

Neuromorphic sensing

Event cameras report asynchronous brightness changes, not fixed-frame video. For navigation, think of them as a way to extend visual motion estimation into conditions where normal cameras fail: microsecond timing cuts latency, and high dynamic range keeps you observing during lighting transitions that would saturate regular sensors.

Limitations come straight from the physics. Static or low-contrast scenes generate few events, so motion becomes hard to constrain. Most pipelines reconstruct pseudo-frames or time surfaces to reuse conventional algorithms, which forces choices about accumulation windows, noise filtering, and latency. Background activity, hot pixels, and flicker can manufacture convincing false motion if you're not careful.

Maturity is spotty. Research looks good in high-dynamics scenarios. Toolchains, calibration, long-duration field tests? Not there yet. Today, event cameras are realistic only as a complement to VIO, not as a primary nav sensor.

Inertial navigation systems

An INS uses gyro and accelerometer measurements to compute attitude, velocity, position without any external signal. It's the continuity backbone in GNSS-denied systems because it doesn't care what the environment looks like. The trade: unbounded drift. INS error growth isn't magic; it's the inevitable result of integrating biases, scale errors, misalignment, and noise into your state.

Gyro bias corrupts attitude, and bad attitude rotates gravity into the horizontal plane. That creates a phantom horizontal acceleration that integrates into velocity and position. Accelerometer bias goes straight to velocity error. Because these mechanisms integrate once or twice, tiny improvements to bias stability can pay huge dividends in position holdover over realistic mission durations.

Sensor grade matters hugely: bias stability and noise differ by orders of magnitude between MEMS, tactical, and navigation-grade. Cheap MEMS work fine with frequent updates but drift fast when unaided and don't like temperature swings or vibration. Tactical-grade extends your unaided window but costs more and consumes more power and size. Navigation-grade lets you go much longer without updates, but you'll pay, you'll pack bigger, and you'll hit supply chain and export restrictions. That last point isn't academic in European programs trying to cut dependence on outside sources.

Whether your lab calibration model actually works in the field is another question. Temperature compensation, scale factor drift, g-sensitivity, vibration rectification errors. These dominate field failures routinely. Shock can inject bias steps that break your random-walk model. Alignment is also a field mode: autonomous systems need to cold-start and recover from resets without human help, which means you need a way to re-establish attitude and bias without intervention.

How you fuse aiding turns predictable drift into something manageable. Loosely coupled systems accept position and velocity updates and stay modular, but they can collapse suddenly when the aiding dies. Tightly coupled filters update directly from raw data (camera features, altimeter profiles, ranges), so you observe better during partial degradation and catch bad updates. Real field failures aren't “INS broke.” They're an estimator that became overconfident and stopped noticing its updates were garbage.

Magnetic navigation

Magnetic navigation matches measured field signatures against stored maps, just like terrain. Without a map, a magnetometer is attitude reference only. Your positioning capability depends on where you are and what's in your map database.

Outdoors, regional fields come from geology and can be stable enough for correlation where structure is distinctive. Indoors, man-made sources dominate: steel, rebar, machinery, electrical infrastructure. Indoor fingerprinting can work in stable environments but breaks when buildings change, when ferromagnetic objects move, or when electrical loads shift.

Interference kills reliability. Your own payloads, motors, and power electronics generate fields you have to calibrate and manage (hard-iron, soft-iron effects), and they shift with configuration. External interference from vehicles, power lines, or intentional field generation can bias your measurements in ways you can't untangle from true anomalies unless you have other observables.

Physics limits it: altitude increases mean anomalies smear and lose distinctiveness. Magnetic nav works better for ground and low-altitude platforms than high-altitude. Realistically, it works best as an occasional absolute update in a fused system, not as a GNSS replacement.

Don't believe accuracy claims without understanding the conditions. Perception-based systems can look good short-term but drift without bound and fail abruptly in ways environment controls: tracking loss, wrong associations. Map correlation can cap error where terrain or anomalies are distinctive, but geography determines whether it works, and stale or incorrectly registered maps kill it.

Environmental sensitivity is different for each method. INS ignores the scene but cares about temperature, vibration, shock. VO needs texture and light, breaks with obscurants and dirty lenses. TRN depends on sensor geometry and surface interaction, fails flat. Magnetic navigation needs electromagnetic quiet and stable anomalies; strong in some places, impossible in others.

Compute requirements differ too. VO/VIO demand heavy front-end and optimization on edge hardware with strict timing. TRN and magnetic correlation move the load to map storage, hypothesis search, and map production logistics. Fusion complexity scales nonlinearly: every new sensor brings calibration parameters, timing paths, and failure modes you have to watch.

Real GNSS-denied systems don't bet on a single modality. Standard pattern: an inertial core, vision-based aiding for local motion when available, and a map-referenced aid for occasional absolute corrections. Top it with integrity monitoring that can reject bad updates and flag when position isn't trustworthy.

Market direction

GNSS vulnerability isn't treated as an outlier anymore in defense or high-consequence commercial systems. Specs are shifting from best-case accuracy to performance under defined outages and deception. What matters: detection capability, bounded error growth, and safe fallback when you can't trust your position.

The problem isn't unique to military. North Sea energy platforms operate in known jamming zones from vessels and land transmitters. Arctic and polar surveys see degraded geometry and ionospheric problems that make GNSS unreliable for weeks. Underground mining, deep-sea work, dense urban port operations all hit satellite gaps from geometry alone. The core engineering problem (bounded position when GNSS fails) is the same across all these, even if the threat vector differs.

Defense spending focuses on assured PNT: measurable uncertainty, integrity checks, proven jamming and spoofing resistance. Terrain and tactical inertial stay in the game where geography and budget allow. Vision gets valued for being passive and hard to jam electronically, provided you actually qualify it for real weather, dirt, platform vibration, not just clean test footage.

European autonomy adds another dimension: control of sensors, timing, and especially geospatial data and update pipelines. “Infrastructure-independent autonomy” isn't just operating without satellites. It's operating without depending on foreign data, external permissions, supply chains you don't control, or trust anchors that disappear when politics change.

Edge compute helps. It doesn't replace good engineering discipline. Power and thermal budgets limit what you can sustain. Safety arguments get harder when performance depends on whether the cpu is busy. Systems that work in the field treat GNSS as one unreliable measurement, fuse with conservative uncertainty budgets, and test failure modes in realistic environments, not marketing test tracks.