Event-based VIO for UAV indoor navigation demonstrates GPS-free precision flight that transfers to terminal guidance

Indoor UAV navigation is a canonical GPS-denied estimation problem. The vehicle needs to know its position to navigate, and no external reference is available. The solutions that work indoors — visual odometry, lidar SLAM, ultrasonic altimeters — have been adapted for outdoor GPS-denied flight, but the indoor regime remains challenging because the environment is dense and high-dynamic-range lighting (windows adjacent to dark rooms) pushes frame cameras to their limits.

A PMC 2025 paper evaluates event-camera VIO specifically for indoor UAV navigation, comparing event-camera-based estimation against a monocular frame-camera baseline and a stereo frame-camera baseline on a suite of indoor sequences including aggressive maneuvers.

The results

On straight-line and gentle-curve trajectories, all three systems produce similar accuracy. The event-camera system starts showing advantages at angular rates above 100 degrees per second, where the frame cameras begin producing blurred images that degrade feature tracking. At 300+ degrees per second — achievable during aggressive indoor maneuvers — the frame-camera systems produce significant drift while the event-camera system maintains centimeter-level accuracy.

The indoor sequences include cases with sudden illumination changes — a UAV flying from a dark corridor into a sunlit atrium — where the frame camera saturates and loses its feature tracks entirely. The event camera, with its per-pixel logarithmic response, handles the illumination change without saturation or event-rate collapse.

The paper also reports on a fast-descent maneuver at 3 meters per second downward velocity, which produces vertical motion blur in frame cameras. The event camera maintains accurate altitude estimation during the descent, which matters for a UAV that needs to land precisely without GPS.

Terminal guidance analogy

The indoor navigation problem and the terminal guidance problem share several properties. Both involve GPS-denied operation. Both include high-angular-rate maneuvers. Both require accurate position estimation relative to a local reference (the room or the target) rather than an absolute global coordinate. Both have illumination conditions that challenge frame cameras — indoor scenes with mixed illumination, outdoor terminal engagements with sun angle variation.

The key difference is timescale. Indoor UAV navigation operates over minutes, with the estimator accumulating drift over hundreds of meters of travel. Terminal guidance operates over seconds, with the estimator accumulating drift over hundreds of meters of range closure. The short duration of the terminal phase actually helps: a VIO system that drifts 1% of distance traveled would accumulate 2 meters of error over 200 meters of range closure, which is within the terminal guidance accuracy budget. A 5% drift would accumulate 10 meters, which exceeds it.

The event-camera VIO system in the paper achieves below 0.5% drift on indoor sequences, which would produce less than 1 meter of error over 200 meters. That’s well within a terminal guidance budget for a proximity fuze rather than a direct-hit requirement. For a direct-hit requirement the absolute error needs to be lower, but the guidance law closes the gap by actively correcting toward the target rather than dead-reckoning.

The combined estimation architecture

A seeker that combines event-camera VIO (for self-navigation) with event-camera target tracking (for relative navigation) has two complementary estimation loops. The VIO loop maintains the seeker’s own position and attitude estimate relative to a local frame defined at seeker activation. The target tracking loop maintains the target’s position estimate relative to the seeker. Together they give the relative geometry needed to close the guidance law, without requiring either GPS or an external navigation aid.

The PMC paper validates the self-navigation half of this architecture in a demanding environment. Combining it with the event-camera target tracking results from Leonardo, DAGr, and the other papers we’ve covered gives a fairly complete picture of what the combined system can achieve — and suggests that the terminal guidance architecture based on event cameras has sufficient performance for the relevant target classes.