Applications / Optical flow · ego-motion · visual navigation

Navigation that does not wait for a frame

Optical flow, ego-motion removal and visual navigation for robots and uncrewed platforms that move fast, work in poor light, or cannot count on GPS.

Demonstrated on field dataShown on recorded field data from moving sensors

Concept illustration: event data smeared by a rotating sensor on the left, the same scene recovered after ego-motion compensation on the right, with optical-flow vectors expanding from the direction of travel — Concept. A moving sensor turns a still scene into motion everywhere. Remove the platform's own motion and what remains is structure, plus whatever is moving on its own.

The problem

Vision fails when the platform needs it most

Fast motion smears frames. Low light stretches exposures and makes it worse. Glare and shadow saturate one part of the image while the rest goes dark. Those are the conditions where a robot most needs its eyes, and where frame-based visual navigation loses lock.

Without GPS, the camera is often the only external reference left.

Why event sensing

Motion is the native signalEvery event is a change at a pixel. Flow comes out of the data directly instead of being reconstructed by matching one frame to the next.

No blur at speedThere is no exposure window to smear across. Edges stay sharp however fast the platform turns.

Remove your own motionCompensate for the platform’s motion and the scene collapses back to structure. Anything still moving is moving on its own, which is the same step that lets a moving sensor spot a target.

Starlight to sunlightEach pixel adapts on its own, so a scene with deep shadow and direct sun stays usable end to end.

Fits a small robotProcessing runs on the sensor module at a few watts, with no GPU and no off-board link.

Evidence

What we have shown

A rotating sensor’s raw events next to the same events after de-rotation, and a satellite track held while the camera was turned by hand.

Left: raw events from a rotating sensor. Right: the same events after ego-motion compensation.

Event-camera view of an ISS pass with the camera turned by hand, the ISS track held against moving trees and stars — Oct 2026Same ISS pass, camera turned by hand to follow it. The scene sweeps; the track holds, because the stars give the pipeline its own motion reference.

Three-camera sensor head: two event cameras flanking a frame camera on a common plate — Sep 2026Dual-modality head: two event cameras flanking a frame camera on a shared timebase, toed out for wider azimuth coverage.

Still frame of de-rotated event data — Sep 2026De-rotated event data.

Own a program in this space?

If size, power, cost or a frame sensor’s limits are what stand between you and the requirement, we would like to hear about it.

Start a conversation See the platform