GPS-Denied Navigation: VIO and Terrain-Relative Nav in Contested Environments

From November 2022 onward, Russian electronic warfare units operating across the Ukrainian theater pushed the GPS L1 noise floor in some sectors above 50 dB-Hz over civilian frequencies.¹ Within the same window, Ukrainian operators were reporting Mavic 3 GNSS lock loss inside a 30 km arc of the front line. By the spring of 2024, GPS denial had been observed across air corridors over the Baltic Sea, the Eastern Mediterranean, and the Persian Gulf — not all of it intentional, none of it accidental in aggregate.²

The lesson for combat UAS designers was not new. It was just expensive. Visual-inertial odometry (VIO), terrain-relative navigation (TRN), and sensor fusion under denied GNSS are not exotic capabilities to be added to a platform — they are the baseline a 2026 platform has to ship with on day one, or it is not a combat-credible system.

[01] · WHY GNSS LOSESThe threat surface in 2026

GNSS denial decomposes into four operational threats, in roughly increasing order of severity:

1. Noise jamming. Wideband RF energy raises the L1/L2 noise floor and degrades signal-to-noise ratio below tracking threshold. Cheap, ubiquitous, defeated only by directional antennas and notch filtering.
2. Deceptive jamming. The jammer transmits a credible-looking GNSS signal with falsified pseudorange data. The receiver locks to it and reports a coherent position that is wrong by hundreds of meters to tens of kilometers.
3. Replay attacks. Recorded GNSS signal from a different location is rebroadcast in the target area. Modern military receivers detect this through SAASM or M-code authentication; commercial receivers do not.
4. Constellation degradation. Long-term concern, not yet operational. Adversary anti-satellite action against the GPS constellation itself.

Against threats (2) and (3), simply hardening the GNSS receiver is insufficient. The system must be able to detect that the GNSS solution is lying — and to keep flying when it does.

[02] · VIO FUNDAMENTALSWhat visual-inertial odometry actually does

Visual-inertial odometry estimates the platform's 6-degree-of-freedom pose by fusing two complementary sensors: an inertial measurement unit (IMU) sampling at high rate (200–1000 Hz) and one or more cameras sampling at lower rate (20–60 Hz).³ The IMU provides short-horizon dead-reckoning. The camera provides absolute reference against visual features in the environment, correcting IMU drift.

The mathematics has been well-understood since the late 2000s, with MSCKF (Mourikis & Roumeliotis, 2007) and the family of optimization-based estimators that followed (OKVIS, VINS-Mono, OpenVINS) defining the algorithmic baseline.⁴ What changed between then and now is not the algorithm — it is the compute envelope. A modern Jetson Orin NX delivers approximately 70 TOPS in INT8 within a 25-watt power budget, which is enough to run a state-of-the-art VIO front-end at 60 Hz alongside an object-detection inference workload on the same chip. In 2015, this required a workstation tethered to a power cart.

The drift problem

VIO error accumulates without bound over time. The question is not whether it drifts — it does — but how slowly. Reported drift performance in the literature varies wildly with operating conditions:

The VertexOS figure is from internal flight test data collected during March 2026 trials at a Southern California test range, under representative wind, lighting, and emissions-controlled conditions. Independent third-party verification is scheduled for the June 2026 NTC demonstration.

[03] · TERRAIN-RELATIVE NAVIGATIONThe on-board prior that anchors VIO

VIO alone diverges. To bound the error, the system needs an absolute reference that does not depend on GNSS. Terrain-relative navigation (TRN) provides it: the platform carries an on-board prior — a digital elevation model, satellite imagery, or a fused visual-terrain database — and matches what its sensors are currently seeing to that prior.

The technique has heritage. NASA's Mars 2020 mission used a TRN system to land Perseverance within meters of a target point in Jezero Crater, using descent imagery matched against a pre-computed orbital map.⁵ Tomahawk cruise missiles have used TERCOM since the 1980s. What is new is doing this on a 4-pound airframe with a $400 compute module.

VertexOS implements two TRN modes:

• EO-to-orthoimagery matching. Downward-looking EO frames are matched against a compressed orthoimagery prior of the operating area, loaded pre-mission. Matches yield absolute geolocation fixes at ~0.5 Hz under good lighting.
• LWIR-to-DEM matching. At night or under heavy smoke, downward LWIR is matched against terrain elevation derived from the DEM. Lower update rate, lower geolocation precision, but functional in EO-denied conditions.

[04] · SENSOR FUSION ARCHITECTUREHow VertexOS stitches it together

The fusion architecture is the part that determines whether all of this works as a combat system or as a benchmark on a research paper. VertexOS uses a tightly-coupled error-state Kalman filter (ESKF) with the following measurement inputs:

• IMU (accel + gyro, 1 kHz)
• Stereo VIO (60 Hz)
• Magnetometer (when not in magnetically-disturbed environments)
• Barometer
• GNSS, when available and authenticated — treated as opportunistic, never as truth
• TRN fixes (asynchronous, 0.2–1 Hz depending on mode)
• RF-silent mode flag (suppresses any active emission)

The critical design choice is the GNSS gating. Most commercial autopilots trust GNSS by default and ignore inertial when the two disagree. VertexOS reverses this: VIO and TRN are the trusted reference, and any GNSS measurement that disagrees with them by more than a residual-test threshold is rejected and flagged as a deception event. The platform continues flying on VIO+TRN+inertial, and the deception event is logged for post-mission analysis.

The autonomy stack is not robust to GPS denial because it ignores GPS. It is robust because it can detect that GPS is lying and keep flying anyway.

[05] · OPERATIONAL IMPLICATIONSWhat this changes for the operator

The shift from GNSS-primary to VIO+TRN-primary navigation has three operational consequences that procurement requirements writers should specify:

1. Mission planning includes terrain priors. The platform requires pre-loaded orthoimagery and DEM for the operating area. This is a logistics burden — the priors must be current, classified appropriately, and pushed to the platform pre-flight. Operators trained on GNSS-only systems do not always know this.
2. Endurance is degraded in feature-poor terrain. Over open water, uniform desert, or thick cloud, VIO drift accumulates faster. TRN works when DEM has relief; it does not work on a featureless plain. Mission planners need to allocate routes accordingly.
3. RF-silent operation becomes possible. An aircraft that does not need GNSS does not need to receive GNSS. Combined with directional, low-probability-of-intercept downlinks, the entire platform can operate as a near-passive emitter, dramatically reducing its electronic signature.

[06] · CLOSINGThe autonomy stack is the air defense

For Vertex Autonomy, GPS-denied navigation is not an add-on capability. It is the foundation on which every Vertex platform — the X-4 Raptor, the X-7 Talon, the X-10 Sentinel — is designed. The VertexOS stack ships as the navigation primary on every airframe and is the same kernel across the portfolio.

For procurement officers, the question to ask is not "does the system support GPS-denied operation." The question is "what is the documented drift performance under operationally-realistic jamming, and what is the failure mode when the visual sensor itself is degraded." If the vendor cannot answer the second question in writing, the system is not combat-credible.

See the VertexOS stack · See the platforms