How a Roomba decides where to clean

It's not random. A modern Roomba runs SLAM, builds a map of your home, and plans a route. The early Roombas (pre-2015) really did bounce around at random — they just covered enough distance to hit most of the floor eventually. The current ones (Roomba j7+, j9+, Roborock S8) are genuine little robots.

Here's what's inside.

The eight things that make a Roomba a robot

A lidar (on premium models) or a camera (on cheaper ones). This is how the Roomba sees the world. Premium Roombas have a small spinning lidar on top, doing 5-10 full sweeps per second. Cheaper models use a forward-facing camera and computer vision.

A bumper sensor. When the Roomba bumps into something its sensors didn't see (a chair leg, a power cable), this senses the contact and tells the brain to back off.

Cliff sensors — infrared detectors under the edges. They notice when the floor suddenly drops away (the top of a stairwell, the edge of a sunken living room) and tell the Roomba to stop.

Wheel encoders. Tiny sensors that count the rotations of each wheel. This is how the Roomba estimates how far it's moved between sensor updates — basic odometry.

Dirt-detect microphone. Yes, really. There's a piezo microphone above the suction inlet. Crumbs and gritty dirt make a different sound from clean floor; the microphone tells the brain to slow down and scrub more carefully when it detects "dirty" sounds.

A brain. A small ARM processor running custom firmware. The premium models have an additional inference chip for running their computer-vision models.

Three brushes plus a vacuum. The actual cleaning hardware.

A battery and dock. When the Roomba's battery gets low, the dock emits an infrared beacon and the Roomba follows the beacon home, using camera or lidar to do the final precise alignment.

How the map gets built

On the first cleaning run, the Roomba doesn't know your home. It uses its lidar or camera to do SLAM — building a map of the rooms while figuring out where it is on that map.

By the end of run 1, the Roomba has a decent map. By run 2 or 3, the map stabilises — the Roomba recognises the same walls and furniture each time.

Once it has a stable map, you can name rooms in the app ("Kitchen", "Living Room"), exclude zones ("don't clean my baby's nursery"), and schedule selective cleans ("kitchen only on Tuesdays").

How the route gets chosen

Different vendors choose different strategies. iRobot's current approach is boustrophedon coverage — that's a fancy word for "back and forth like a lawnmower". The robot picks a direction, drives in straight lines, then reverses 180° and offsets by a robot-width. This is provably efficient for rectangular areas.

For non-rectangular rooms, the algorithm decomposes the room into roughly rectangular sub-areas, covers each, then moves to the next. Edges and corners get a final pass with the side brush.

How it was built

The premium-tier Roomba is a small marvel of low-cost robotics: $700-900 retail for everything described above. The lidar alone would have cost $10,000+ on a research robot in 2012. The economics work because vacuum cleaners are a multi-billion-dollar market — there's volume to amortise the engineering over.

What's still wrong with it

Wires confuse it. Black furniture confuses the camera (low contrast). Stairs are still risky. Pet accidents are catastrophic (the j7+ does have a feature to avoid pet waste using its camera; iRobot reportedly guarantees a refund if it fails). And the suction is real but limited compared to a corded upright vacuum.

---

Curious how the same SLAM technique works on a Mars rover or a humanoid? Read SLAM.