How a robot mower decides where to go

Most robot mowers don't know where they are. Here's what that costs, and what changes when they do. Play with the lawn below.

Watch the first robot mower you ever see and it looks drunk. It drives in a straight line, hits the edge of the lawn, turns a random amount, and sets off again. No pattern. No memory. It will happily cut the same patch five times while a strip by the fence stays long for a week.

That isn’t a bug. For twenty years it was the design.

Why random works at all

The early mowers had one piece of information: a buried wire around the lawn. Cross it and a sensor says “you’ve left”. That’s enough to bounce around inside a shape, and if you bounce for long enough you cover all of it. Run it every day and the lawn is never long enough for anyone to notice the gaps.

Try the random mode above at 16× and watch the counter. It gets to half the lawn fast. Ninety percent takes noticeably longer. The last few percent are where it spends most of its life, because by then it’s mostly re-cutting grass it already did. The “wasted passes” number is the share of its work that was repeat work.

Battery, blade wear and noise all scale with that number.

What changes when it knows where it is

Switch to lines. The mower drives in stripes one mower-width apart, finishes, and stops. Same lawn, a fraction of the time, almost no repeat work, and it looks like a lawn someone cared about.

The catch is the word knows. To cut in stripes a mower has to know its position to within a few centimetres, constantly, under trees and next to walls. Ordinary GPS is out by metres. The answer on the newer machines is a mix of:

• RTK GPS, a base station in the garden that corrects the satellite signal down to about two centimetres, as long as the sky is open.
• Cameras and vision, so the mower can recognise the edge of the grass itself and tell a lawn from a path, a flower bed or a cat.
• Wheel odometry and an IMU, so it can keep a rough fix for a few seconds when the other two drop out.

Fuse those together and you get a mower that builds a map on the first run and plans stripes on every run after that. No wire. The expensive part isn’t the motor, it’s the software deciding which sensor to trust right now.

What the toy doesn’t model

The simulation above is flat, dry and rectangular. Real lawns have slopes that make wheels slip, wet grass that clogs blades, and shadows that confuse cameras. A mower that stripes beautifully in June can lose its position under a wet October canopy and fall back to bouncing. Most of the interesting engineering is in that fallback, not in the happy path.

Why we care

Kestrel is a hedge trimmer that hovers. A hedge is a lawn turned on its side, with a harder localisation problem and a sharper penalty for getting it wrong. Everything above about knowing where you are applies, except the mower can’t rest on its wheels while it thinks.

That’s the problem we’re interested in. These notes are where we work it out in public.