How a robot vacuum decides where it has already been

Everything below about robot vacuum navigation comes from what actually happens rather than from what is supposed to.

The cheap approach is statistical

Early and budget robots move in straight lines until they hit something, turn through a pseudo-random angle and continue. Given enough time this covers most of a room, which is why the strategy works at all.

It is inefficient, repeats areas, misses others, and cannot report what it did. It also requires almost no sensing, which is exactly why it is cheap.

Mapping robots solve two problems at once

Simultaneous localisation and mapping means building a map of an unknown space while simultaneously working out where you are within it. These are interdependent: you need a map to locate yourself and your location to build the map, so the algorithms estimate both together. Spinning laser rangefinders measure distance in all directions many times a second, which makes the estimate tractable.

In the datasheet, wheel rotation supplies the short-term motion estimate that the ranging data then corrects, which is why deep carpet and slippery floors both degrade mapping — one lets the wheels slip and the other lets them spin.

Cameras and lasers fail differently

Laser-based navigation works in complete darkness and struggles with glass, mirrors and very dark absorbing surfaces. Camera-based navigation is cheaper and needs light, and it can recognise objects, which is how obstacle avoidance for cables and pet mess works.

Higher-end machines combine both, because the failure modes do not overlap. Structured-light and time-of-flight sensors sit between the two, projecting their own illumination so they work in the dark while still being defeated by the same transparent and mirrored surfaces.

Getting lost is the real engineering problem

A robot picked up and moved, or one that slips on a rug, loses its position estimate and must relocalise against its stored map. Doing that reliably in a room that has changed — furniture moved, doors closed — is considerably harder than the original mapping. Most user complaints about erratic behaviour are relocalisation failures rather than mapping failures.

Returning the machine to its dock before restarting it hands the estimator a known starting position, which resolves most cases where a robot has confidently begun cleaning the wrong room.

The map is data about your home

Floor plans, room dimensions and furniture positions are detailed information, and on cloud-connected models it leaves the house. Privacy policies vary substantially and some devices offer local-only operation. This is worth checking specifically, because the map is more revealing than most people assume.

Machines that recognise obstacles may also upload photographs to improve that recognition, and images taken at floor level inside a home are a different category of data from a floor plan.

Implementations differ, and vendors are not obliged to document the differences.

Coverage is decided by the planner, not the sensor

Once a map exists the robot still has to choose an order of rooms, a direction for its passes, and a policy for doorways, rugs and chair legs. Long parallel passes with turns at the ends cover open floor efficiently and cope badly with clutter, where the pattern fragments into short segments and the machine spends its time turning. Virtual barriers are useful precisely because the planner is the weak component: excluding a difficult area is usually cheaper than expecting the software to solve it.

Under load, edges and corners are reached by a side brush that flicks debris into the path of the main brush, so a machine that appears to be missing corners is often failing at the brush rather than at the map.

The takeaway

The hard part is not drawing the map. It is knowing where you are when the map is wrong.

Once you know what it is trading away, the design stops looking arbitrary.