How can a drone navigate without GNSS?

Movement is easier to measure than location

Suppose a drone estimates that it has travelled 100 metres east from its take-off point. That is a position in a local coordinate system. It only becomes a geographic position if the aircraft also knows where that starting point is and how its local axes align with Earth.

This distinction explains why a drone can fly around a building without GNSS yet struggle to follow a long geographic route. Local observations constrain movement around the building. Over a longer flight, errors in estimated speed and heading accumulate unless another reference corrects them.

Stabilisation is a separate task. The flight controller can use inertial measurements to control attitude even after it loses a reliable position estimate. A level aircraft may still be drifting away from its intended path.

Inertial sensors keep track between observations

An inertial measurement unit, or IMU, contains gyroscopes and accelerometers. The navigation software uses them to estimate orientation and changes in velocity. It accounts for gravity, then integrates motion over time to estimate position. These measurements arrive rapidly and do not require a view of the ground.

Small biases are the problem. In a simplified one-axis model, a constant acceleration error of 0.01 m/s² produces 450 metres of position error after five minutes. The error grows with the square of elapsed time.

Real inertial error also depends on gyro bias, vibration and alignment. Other sensors help by giving the estimator observations against which it can check its prediction.

Cameras and LiDAR constrain drift

Visual odometry tracks features across successive images and estimates how the camera moved. Combining those observations with an IMU is called visual-inertial odometry. The IMU predicts motion between frames, while the images help correct the accumulated error.

SLAM means simultaneous localisation and mapping. It estimates the aircraft's position while building or updating a map. Recognising a place already visited can help correct earlier drift. LiDAR adds direct measurements of surrounding geometry, which can help when images contain little useful texture. The FAST-LIVO2 research implementation combines LiDAR, camera and inertial measurements in one estimator.

None of these observations is equally useful everywhere. Motion blur can erase image features. Snow can hide familiar ground patterns. Repeated roofs or rows of identical buildings can produce the wrong match. LiDAR also needs enough distinct geometry to constrain movement. Adding sensors helps when they supply information that the others are missing.

A map connects the local estimate to Earth

A georeferenced map gives known coordinates to observed features. Software can compare an aerial image with satellite imagery, or match measured geometry to a surveyed structure. If the match is correct, it supplies a geographic correction to the local motion estimate.

One plausible image match is weak evidence. A sequence is more useful because the candidate locations must agree with the aircraft's movement. If one match implies a jump that the aircraft could not have made, the estimator has a reason to reject it.

AI can help recognise places and find features that remain identifiable across different views. The navigation software still has to check those matches against geometry and motion. A recognition score measures confidence in the match, not the distance by which the aircraft's position could be wrong.

What the published results show

NaviLoc, matching a flight trajectory to satellite imagery

Shpagin and Panchenko's NaviLoc paper combines image-to-map matches with relative motion estimates. On its low-altitude rural benchmark, the authors reported a mean localisation error of 19.5 metres. This measures performance on the evaluated trajectory, rather than a guaranteed error bound for a new route. The paper also restricts access to its evaluation data and code, which limits independent reproduction.

AIRSENSE, image-based navigation in flight

The European Defence Agency reported an onboard image-based navigation demonstration over land during the September 2026 REPMUS exercise in Portugal. The aircraft flew more than 100 kilometres from its ground station. That is the reported separation from the station, not the length of a documented uninterrupted GNSS-denied flight segment.

Known landmarks, with errors still measured in metres

A 2026 study by Zheng and colleagues used turbine towers as visual landmarks. Field validation at a 26-turbine wind farm produced a three-dimensional root-mean-square error of 24.95 metres, with about 220 milliseconds per processed frame on a Jetson Xavier NX. That error is too large to treat the result as a precision-landing solution.

These studies answer different questions. Their error figures use different metrics and test conditions, so they do not form a direct ranking of navigation systems. They do show why geographic reacquisition and close-range flight need separate performance requirements.

Sensor fusion depends on timing and calibration

The estimator must know when each measurement was taken and where each sensor sits on the aircraft. A delayed camera observation describes an earlier position. Treating it as current can make the measurements disagree even when the camera and IMU are working correctly.

The PX4 navigation filter documentation describes how an autopilot combines measurements and checks them against predicted motion. An external vision system can supply position estimates and their uncertainty. Integration still requires consistent coordinate axes, sensor offsets and timestamps.

More accurate average position is only part of the job. The aircraft also needs a timely warning when the estimate is unreliable. An error that the system detects before a turn has a different consequence from the same error discovered after the aircraft has crossed an obstacle boundary.

Test the failures along the intended route

Begin with recorded sensor data and an independent position reference. Compare actual error with the uncertainty reported by the estimator. Include the terrain, lighting and flight motions expected on the route, then test what happens when individual observations become unavailable or wrong.

The camera loses features

Blur or obscure recorded images. Measure how long position error takes to exceed the route's clearance limit and when the estimator reports the loss.

The map match is wrong

Supply a plausible match to a different location. Check whether motion and range measurements reject it before it changes the commanded path.

Sensor timing is wrong

Introduce a known delay between camera and IMU data. Compare the reported uncertainty with the error measured by an independent reference.

A position source fails

Remove GNSS or companion-computer updates in simulation and hardware-in-the-loop tests. Verify the aircraft's response with the remaining sensors.

Choose the response to lost positioning before flight. Hovering requires a usable position estimate; returning home requires a way to locate home. The available response depends on the aircraft, remaining sensors and landing options. Measure the time available to act as well as the time needed to detect the failure.

Flight approval is a separate requirement. In Norway, many operations beyond visual line of sight fall in the Specific category. The Civil Aviation Authority Norway describes the available authorisation routes. For an operator evaluating adoption, the costs also include reference maps, calibration, sensor upkeep and crew time, not just the navigation computer.

Sources

Reviewed 1 November 2026. The results above come from the cited research teams. The drift calculation and suggested failure tests are explanatory examples, not Lyon Industries flight-test results.

1. European Defence Agency, AIRSENSE flight demonstration

   16 October 2026. Image-based navigation trial during REPMUS in Portugal.

2. Shpagin and Panchenko, NaviLoc

   Drones, January 2026. Aerial-to-satellite localisation using image matches and relative motion.

3. Zheng and colleagues, landmark-aware visual localisation

   Scientific Reports, September 2026. Turbine-based localisation tested at a 26-turbine wind farm.

4. HKU-MARS, FAST-LIVO2

   Research code and datasets for LiDAR-inertial-visual odometry.

5. PX4, navigation filter documentation

   How the autopilot combines measurements, handles timing and checks observations against its estimates.

6. Civil Aviation Authority Norway, Specific category

   Authorisation routes for operations including flights beyond visual line of sight.