visual-inertial odometry

Visual-inertial odometry, or VIO, is a way for a robot to track its own motion by blending two very different senses: what its camera sees and what its inner sense of motion feels. The camera watches the world slide past — as you walk forward, nearby objects sweep by quickly while distant ones barely shift, and from that flow the robot can read how it has moved. The inertial sensor, meanwhile, is like the feeling in your gut when an elevator starts: it senses acceleration and turning directly, with no need to look at anything. VIO marries these two so the robot can estimate where it has gone using just a small camera and a tiny motion chip, no GPS required.

Each sense covers the other's weakness, which is the whole point. A camera gives a rich, detailed read of motion but stumbles when the view is blurry, dark, or staring at a blank wall with nothing to track; it is also relatively slow, delivering only tens of frames a second. The inertial sensor is lightning-fast and works in pitch darkness, but on its own it drifts badly because it must add up its readings and the small errors snowball within seconds. Fused together, the fast inertial sensor carries the estimate smoothly through the gaps and the blurry moments, while the camera repeatedly anchors it back to what is really there, holding the drift in check.

This is worth contrasting with plain wheel odometry, which estimates motion by counting wheel turns. Wheel odometry is simple and cheap, but it is fooled by anything the wheels do that the body does not — slipping on ice, spinning in mud, or skidding — and it only works for things that roll on wheels. Visual-inertial odometry needs no wheels at all, so it works for drones flying, phones held in your hand, and headsets on your face, and it is not lied to by a slipping wheel; the trade is that it demands more computing power and a scene with enough visible texture to track.