Validating GPS-Denied Navigation: Measuring Drift You Cannot See

"GPS-Denied Capable" Is Not a Specification

The phrase appears on nearly every tactical UAV datasheet now. It is close to meaningless without three accompanying numbers: drift rate, the duration over which that rate holds, and the conditions under which it was measured. A platform that drifts 4 m/min over featureless terrain and one that drifts 40 m/min are both "GPS-denied capable." Only one completes a 20-minute ingress.

The difficulty is that drift is invisible in flight. The aircraft believes its estimated position. So does the operator, because the map display shows the estimate. Without independent ground truth you are watching a system report confidence in a number that is quietly diverging from reality.

Why the Navigation Source Determines the Test

Pure inertial dead reckoning drifts as roughly t², because accelerometer bias integrates twice into position. A MEMS IMU with 0.05 m/s² residual bias accumulates about 90 m of position error in 60 seconds. This is why nothing serious relies on inertial alone beyond very short intervals.

Visual-inertial odometry drifts approximately linearly with distance travelled, typically 0.5-2% of path length in good conditions — but it fails discontinuously rather than gradually. Over water, fresh snow, uniform desert, or dense smoke, feature tracking loses lock and the estimate reverts to inertial with all of its t² behaviour. The failure is not graceful degradation; it is a cliff.

Terrain-referenced navigation is stable over its reference area and useless outside it. Each of these needs a different test, and a campaign that treats "GPS-denied" as one capability will characterise none of them.

Field Example: Excellent Until the Reservoir

We tested a VIO-equipped platform that performed genuinely well — 0.8% of path length over mixed agricultural terrain, comfortably inside spec. The programme was ready to sign off.

We added a leg crossing a 1.4 km reservoir. Feature lock was lost 200 m in. The estimator fell back to inertial and by the far shore the reported position was 310 m from truth. Worse, on reacquiring features over land the filter did not reject the accumulated error; it fused the bad prior and took a further 90 seconds to converge, during which the aircraft flew a slow curving correction that would have been unacceptable on any real ingress. The capability was real. The characterisation was incomplete, and the incomplete part was exactly the terrain the customer intended to cross.

Building a Test That Produces a Number

Establish independent ground truth — RTK logging onboard, recorded but not fused, is the cleanest method. Denial must be genuine: disable the receiver in firmware rather than relying on a jammer, so you know precisely what the estimator had available.

Fly each terrain class separately and report them separately. Include transitions between classes, because that is where estimators fail worst. Extend runs to at least twice the intended mission duration; drift is rarely linear and the interesting behaviour lives in the tail. Then publish drift as metres per minute per terrain class, with the reacquisition transient stated explicitly. That is a number an integrator can plan a mission around.