Asked how a window cleaning robot works, the answer is short: vacuum suction grips the glass, belts or tracks drive it, a squeegee wipes water into a recovery tank. Vacuum pump, drive, and edge sensor decide whether it stays stuck.
That is the whole machine. No magic, no AI, no lidar. The interesting engineering is entirely about keeping a 5 kg box attached to a vertical or tilted surface while water is being sprayed around it.
How does the suction system keep a robot on vertical glass?
A vacuum pump inside the chassis pulls air out of the cavity under the unit, creating a pressure difference against atmosphere. Most 2026 machines hold 2000-4000 Pa. Sea-level atmosphere is about 101,325 Pa, so the unit is running at a pressure deficit of roughly 2-4 percent — enough to hold several times its own weight on dry glass, less on wet.
Water is the enemy of that math. A squeegee pushes a film ahead of the machine, and if the film reaches the suction chamber, pressure equalises through the liquid. Good machines isolate the vacuum cavity from the wipe path with a rubber skirt. Cheap ones do not, and this is why budget units drop on humid days.
Two suction architectures are common. Single large cavity: strong hold, but loses everything if the seal breaks at one point. Multi-cup array: partial loss, remaining cups hold, and the machine can re-seat itself. Our Lingyun Y3 review covers a four-cup layout and what happens when one cup passes over a seal bead.
What actually moves the robot across the glass?
Four approaches, in order of how often we see them on commercial skylights:
| Drive type | How it moves | Grip on wet glass | Weak point |
|---|---|---|---|
| Wheel + belt drive | Rubber wheels with independent suction | Good to 45 deg | Belts clog with grit |
| Tank tracks | Continuous track on glass | Excellent on steep | Heavy, marks on soft coatings |
| Squeegee glide | Oscillating pad, no wheels | Poor above 25 deg | Slow, 8-12 m2/h |
| Magnetic (steel frames only) | Magnet rollers | Excellent | Requires steel, not aluminium |
Speed matters for planning, not for bragging. A typical commercial unit covers 45-70 m2/h in a straight field. Add rafter crossings and the number halves. We usually quote 30 m2/h in a real roof estimate, and we have never been wrong in the optimistic direction.
How does the robot know where the roof edge is?
Older units used a bumper switch — the machine hit the frame, stopped, reversed 150 mm. That works but wastes a strip of glass 40 mm wide. Current machines use three sensing methods, sometimes together:
- Infrared reflectance. An IR emitter bounces light off the glass underside; a step or frame changes the return. Good at 5-15 mm range.
- Ultrasonic. Works through dirt film and slight scratches. Slower response, 20-40 ms.
- Encoder dead reckoning. Tracks revolutions and estimates position. Drift of 3-5 percent per pass is normal, which is why mapped routes still need sensors.
If a vendor only mentions one of these in the datasheet, ask which one handles the edge. Edge detection is the single most common cause of machines falling or stalling, and it is a spec most buyers never read.
Where does the dirty water go?
Two-tank or one-tank. Two-tank designs spray clean water and recover dirty water separately, which keeps the pad from redistributing grime. Recovery cylinders hold 0.4-0.8 L on commercial skylight units. One-tank machines just cycle the same water and leave a haze.
Practical consequence: a 180 m2 glass roof produces roughly 4-6 L of dirty water, and a 0.6 L recovery tank means eight to ten stops. Plan the route around the tank, not around the roof shape. It is the same advice we give in the Lingfeng S1 review for stepped roofs.
Does the machine need a person on the roof?
Yes. Every unit we know of in 2026 needs someone to place it, tether it, refill it, and lift it over a rafter. Some commercial systems run on rails across a solar array and can work unattended, but those are installed infrastructure, not portable devices. Marketing language about “fully autonomous facade cleaning” usually means the machine is autonomous between refills, which is a much narrower claim.
That person still needs fall protection under EN 363 or the equivalent local rule. The robot removes the requirement to be near the edge with a pole or a harness line — it does not remove the requirement to be on the roof safely.
One design detail to look at if you buy: does the unit have a physical bypass valve that opens when the vacuum chamber pressure drops? Without one, the pump runs at full load trying to recover a leak, and pump life drops by roughly a third on dirty sites. It is a EUR 15 part that separates good machines from adequate ones.
Key Takeaways
- Suction, drive, and edge sensing are the three subsystems that decide reliability.
- 2026 machines hold 2000-4000 Pa; wet glass cuts effective grip by 20-40 percent.
- Real throughput is about 30 m2/h on a framed skylight field, not the 70 m2/h in brochures.
- Two-tank water recovery prevents haze; 0.6 L tanks mean 8-10 refills on a large roof.
- A human still places, tethers, and refills the unit. “Autonomous” means between refills.