Skylight cleaning robots in 2026 navigate with a mix of downward optical sensors, edge-detection probes, IMU heading data and pre-taught waypoints. They do not use GPS, which is unreliable within 2 metres of a large building anyway.
Navigation is the least glamorous part of the machine and the part that decides whether it survives its first month on the roof. Get it wrong and you get a robot parked at the bottom of an atrium.
Why is GPS useless on a glass roof?
Two reasons. First, a large building masks satellite signals, so position drifts by several metres. Second, glass offers no visual texture for conventional camera tracking, and a mirrored surface confuses anything trying to match features. Machinery that cleans solar arrays hits the same wall; the industry has spent a decade solving it with local sensing, and skylight robots borrow that approach.
What sensors does a skylight robot actually use?
- Edge probes: contact or optical detectors that look for the drop beyond the glass. This is the safety-critical layer.
- Downward optical flow: tracks frame lines, mullions and drainage channels to estimate movement. Works well, fails on featureless tinted glass.
- IMU and encoders: measure heading change and wheel or track rotation. Drift compounds over a large roof.
- Pre-taught paths: an operator drives the first pass manually, the machine repeats it. Simple and reliable.
- Optional LiDAR: used on larger facade machines, less common on light skylight units because of weight.
Where does navigation fail in real buildings?
Drainage channels and smooth dark frames are the top troublemaker. If the machine’s camera sees uniform dark glass with no contrast, optical flow stalls and the robot thinks it has not moved. It then over-drives its correction and wanders. A Lingkong K3 handles this with encoder fallback, but slow it down and expect longer cycle times.
Second failure: raised gaskets and panel height steps over 6 mm. The machine catches, corrects, catches again. Third: direct sun causing flare in optical sensors during late morning. Sites in hot climates learn to clean before 9 am or after 4 pm.
Should you programme the route or let the robot learn it?
For a roof you clean four times a year, teach the path once and repeat it. For a large multi-zone roof, mapping pays off because you can split it into named zones and clean the worst one first. Do not over-automate a simple roof; a saved waypoint file and a checklist beat a full mapping project every time on a 600 m2 sawtooth roof.
How do you teach a route on a mirror-like roof?
Slowly, and with marks. On a roof with little visual contrast, the most reliable method is a manual teaching pass at reduced speed, with the operator tapping a waypoint each time the machine crosses a mullion or frame line. Those waypoints anchor the route, and the machine interpolates between them. Without them, on pure tinted glass, position drift can reach a metre over a 30 m run.
Adding a few temporary markers, such as removable tape lines the machine can see, is unglamorous but effective. Many crews keep a set for the first two cleans of a new roof and remove them once the waypoint file is stable.
What is the single biggest cause of a robot getting stuck?
Panel height differences. Anything over about 6 mm is a step the machine has to climb or avoid. Old skylights and mixed replacement panels produce exactly these steps, often invisible in photos because they look flat. Before buying, run a straight edge across the roof and note every step and every raised gasket. That survey will tell you more about whether the machine will work than any datasheet.
| Obstacle | Typical height | Robot response |
|---|---|---|
| Flush frame | 0-3 mm | No issue |
| Raised gasket | 4-8 mm | Slows, may need a re-seat |
| Panel step | Over 8 mm | Often blocked; re-plan the route |
| Drainage channel | 10-20 mm wide | Tracked, but slows optical flow |
Do bigger machines navigate better?
Not automatically. A heavier unit carries more sensors and a longer track base, which helps heading stability, but it also loads the glass more and is harder to reposition at an edge. For skylight work the sweet spot in 2026 is a light machine with good edge sensing rather than a heavy facade unit. Weight is the enemy of navigation on fragile roof glazing.
How much does a taught route drift over time?
Little, if the roof does not change. Mullion positions are fixed, so a waypoint file written in spring is still accurate in autumn. What does drift is the machine itself: worn cups, a slightly different battery, a re-seated wheel. Those small changes add up, and a route that ran cleanly in March may need one re-teach by September. Build a re-teach into your annual maintenance rather than waiting for a visible failure.
The other thing that changes is the roof. New plant, a fresh antenna, a repair to one pane: any of these can block a path that used to be open. Re-survey after any work on the roof, and update the waypoints before the next clean.
Key Takeaways
- Skylight robots use edge probes, optical flow, IMU and taught paths, not GPS.
- Featureless dark glass is the hardest surface for optical tracking.
- Panel steps over 6 mm and strong morning sun cause repeatable navigation faults.
- Teach-and-repeat beats full mapping on small or simple roofs.
- Test navigation on your actual glass before signing. Start at our contact page.