A window cleaning robot works by pulling a low vacuum between its sealing ring and the glass, usually 2,800–4,000 Pa, which is enough to hold 3–6 kg against a vertical pane. Small drive wheels shift it in a Z or N pattern while a microfibre pad and a spray nozzle do the scrubbing. When the sensor at the edge sees glass ending, it reverses.
That is the honest version. Nothing in the machine is exotic. What separates a machine that survives a rooftop gantry in Shanghai from one that drops off your second-floor bedroom window is cheap parts executed well: the seal ring, the pump duty cycle, the bumper sensor and the firmware that decides to stop.
What actually holds the robot on the glass?
Suction, and it is a closed loop. The pump evacuates air from the space between a rubber lip and the pane; atmospheric pressure pushes the chassis back. Lose the seal and the machine loses everything in about a second. On a horizontal skylight the same physics still apply but gravity now works against the pad contact instead of the grip, which is why flat-panel robots need a different wheel compound than the ones built for curtain wall.
Numbers matter here. A 1.2 kg indoor unit like the ECOVACS WINBOT mini 2 runs around 2,800 Pa and is designed for panes up to roughly 3 m². Industrial skylight units such as the Lingdu Intelligence Lingkong K3 sit higher, because a 3 m by 4 m atrium panel with a wet pad can present real peel load. Our breakdown of the Lingkong K3 review walks through what that margin buys you in practice.
How does the robot know where the glass ends?
Three sensing layers usually run at once. A current spike in the pump tells the controller the seal has broken or the pad has hit a frame. A mechanical bumper on the front edge trips on mullions. And an optical or gyro edge check watches for the last 30–50 mm of pane. Firmware then chooses: back off 40 mm, rotate, continue.
Where people get hurt is trusting the third layer alone. Glass with a frosted band, heavy mineral staining, or a deep silicone joint can look like an edge or hide one. If you clean a lot of skylights with factory-applied ceramic frit, ask the supplier for the minimum contrast the optical sensor tolerates. Most will not have a number. That silence is a useful signal.
What happens along the water path?
Water is either sprayed ahead of the pad or fed through it. Spray-first wins on bird lime and dust film; pad-fed wins on water marks. Two tanks is the norm on commercial units: a clean tank around 0.6–1.0 L and a dirty tank that pulls spent water back so it does not streak. On a sun-baked glass roof, the spray volume matters more than the pad quality — 15 ml per m² evaporates before the squeegee passes and leaves a haze you will see from the atrium floor.
Do window cleaning robots really work on vertical glass?
Yes, with conditions. They work well on framed panes up to roughly 2.5 m tall, on glass free of heavy oil, and inside temperature windows that keep the seal rubber soft. They struggle on frameless butt-glazed curtain wall (nowhere to register edges), on panes above 35 °C where the lip goes slick, and on anything with a bird spike or cable running across it.
So who should not buy one? A facilities team with four or five irregularly shaped skylights and no dedicated operator. Setup time, safety line rigging and pad changes will eat the savings. The machines pay off from around 600–800 m² of glass that gets cleaned four or more times a year. Below that, a pole and a good squeegee still win on total cost.
What does the machine leave out?
Judgement. A robot will happily scrub a pane with a hairline crack. It will ride over a loose gasket and push water into the frame cavity. It will not notice that the skylight above the server room has a sealant failure. Anyone running these on a commercial roof needs a two-minute visual before each cycle, and a written rule about stopping when wind gusts pass 8 m/s on an exposed roof.
Runtime is the other quiet limit. A cordless unit rated 90 minutes on dry glass gets roughly 55–65 minutes with the pump working and the spray running, which on a 25 m² skylight cluster is about one full cycle and a half. Read our cordless runtime test notes before you size a battery plan.
How much glass does one charge cover?
Coverage is a multiplication, not a number on a box. Take the runtime, subtract 15% for repositioning, and divide by the pace figure. A machine doing 22 m²/h on clean glass will do about 14 m²/h on a dirty skylight, because the first pass moves quickly and the second pass is where the time goes. On a 12 m² bay that is 50 minutes, not 33.
Two-pass cleaning is not optional on glass that faces the sky. Dust settles on both sides of nothing — the top surface collects everything, and a single pass redistributes it into a film that dries into a haze. The water volume per square metre matters as much as the pad: below about 12 ml per m² on hot glass the spray evaporates before the pad arrives, and you are essentially dry-scrubbing baked-on mineral dust.
That is also the answer to the question people ask after they buy: why does it look worse than hand cleaning? Usually because the machine is doing one pass with too little water, or the pad has residue from the previous cycle. Both are fixable in a day and neither is a machine fault.
Key Takeaways
- Grip comes from a low-vacuum seal, typically 2,800–4,000 Pa; lose the lip and the robot falls within about a second.
- Three sensing layers — pump current, bumper, optical edge — matter more than the marketing spec sheet.
- Spray-first cleaning handles bird lime; pad-fed cleaning leaves fewer water marks on hot glass.
- Robots pay back above roughly 600–800 m² of glass cleaned four times a year, not on a handful of windows.
- Always require a stated minimum sensor contrast and a wind cutoff before signing a purchase order.
If you want the shortlist version of this whole category, our skylight robot reviews hub collects the models we have actually run, including the ones we would not put on a roof.