Ask three suppliers whether their robot navigates with LiDAR or cameras and you will get three confident answers that do not agree. That is because the question is framed wrong. LiDAR vs camera navigation is not a contest with a winner; it is a question about which failure mode your building will actually present — glass and long blank corridors, or crowds and clutter — and whether the machine you are buying has an answer for it.
This page compares the two sensor families honestly: what each one measures, the specific conditions where each one gets into trouble, why production robots carry both, and the five questions that reveal more about a supplier’s navigation than the words “laser” or “visual” ever will. For the wider picture — how sensing, mapping, planning and cleaning fit together — start with the navigation technology explainer.
The question behind every LiDAR vs camera navigation decision
Nobody buying a floor robot cares about sensors. What they care about is whether the machine will still be cleaning the route it was given in week forty, in a building where displays moved, a pallet was left in an aisle, and the evening sun comes through the entrance at a low angle for two months a year.
So read every sensor claim as a claim about reliability in your conditions. “LiDAR” is shorthand for a machine that is strong where geometry is rich and stable. “Cameras” is shorthand for a machine that is strong where objects matter more than distances. Neither shorthand tells you where the machine breaks.
What each sensor family actually measures
| Laser scanning (LiDAR) | Cameras (depth / RGB) | |
|---|---|---|
| Measures | Distance to surfaces, precisely, as a 2D or 3D scan | Shape, colour, texture and movement — what a thing is, not only where it is |
| Works well in | Feature-rich interiors, darkness, long straight runs, stable geometry | Cluttered, crowded, human-filled spaces; recognising objects worth going around |
| Struggles with | Glass and mirrors, reflective or wet floors, very long feature-poor halls, blank walls | Low or changing light, glare, dust on the lens, and surfaces with no visual texture |
| Typical failure looks like | Position drift, lost localisation, phantom obstacles that stop the machine | Missed low or transparent hazards, or a route that wanders when lighting changes |
The two columns are complementary by construction: one is good at where, the other at what. That is the whole reason the debate never resolves.
Where a LiDAR-only machine gets into trouble
Laser navigation needs something to measure. Give it nothing to measure and it drifts.
- Glass and reflective surfaces. Atriums, glazed shopfronts, stainless steel panels and mirrors either absorb the beam or bounce it somewhere unhelpful — the machine reads a wall where there is a doorway.
- Long, feature-poor halls. A 60-metre exhibition hall with smooth walls offers the scan very few landmarks to match against, so small errors accumulate over the length of the run.
- Wet or polished floors. Specular reflections confuse the return signal exactly where a scrubber spends its day.
- Everything at ankle height. A 2D scan gives one slice of the world. Cables, pallet corners and a dropped box may sit below it.
The standard fix is physical rather than electronic: fixed floor markers give the machine known references in exactly the spaces where sensing is thin. It is an inexpensive way to make navigation dependable in a hard building, and it is why markers appear in project lists far more often than they appear in brochures.
Where a camera-only machine gets into trouble
Cameras bring the opposite weaknesses, and they are the ones that show up on a live site rather than in a demo:
- Lighting that changes. Evening light through a glazed entrance, fluorescent flicker, a section of the floor where the lamps are off — all of them change what the camera interprets.
- Blank and repetitive surfaces. A long wall of identical white panels or a plain warehouse bay gives visual matching very little to lock onto.
- Occlusion. In a crowd, people block the view of what is behind them; a single camera can lose the context that a 3D scan keeps.
- Wear and maintenance. A dusty or smeared lens degrades quietly, and the operator often notices the symptom (a wandering route) long before the cause.
None of these are arguments against cameras — they are arguments against cameras alone, in buildings that have glass, corridors or crowds.
Why production machines fuse both
The robots that survive real buildings do not pick a side. They resolve the ambiguity between sensors and add two protections that neither family provides:
- Sensor fusion. The scan gives structure, the cameras give meaning, and ultrasonic sensing catches the low, transparent and reflective things optical sensors misread. When one source is weak, the others carry the position estimate.
- Drop-off detection. Stairs, dock edges and mezzanine openings are the one mistake there is no recovering from, so they get their own sensing rather than relying on the general navigation stack.
- Contact bumpers. The honest admission that sensing is not perfect: if something is too soft or too low to be seen, the machine stops on touch instead of driving through it.
Two of our machines can be described concretely rather than in adjectives. The SC80 runs multi-line and single-line LiDAR together with six ultrasonic sensors and three cameras. The Scrubber 50 pairs 2D LiDAR with a 3D depth camera, an RGB camera and collision sensors. The full pipeline — how those readings become a map, a position and a route — is set out step by step in the navigation explainer.
Adaptability matters as much as the sensor list. On the waterless SW55A, routes are re-planned when the layout changes rather than hard-coded to a fixed line, so a moved display changes the path instead of breaking the task. The SW80A maps up to 40,000 m² in a single survey, which is what lets a large site absorb change without a re-mapping exercise every time something moves.
Two newer units show the same pattern at different price points. The C5 positions on laser and vision together — dual positioning, in the manufacturer’s own description — and fuses point cloud with vision for obstacle recognition, which is what lets it tell a pallet from a person at range. The SW GT layers 3D LiDAR, RGB-D vision and ultrasonic sensing into 720° coverage, and is specified for suspended, low-lying and moving obstacles: the three categories where a single sensor family fails.
Three claims worth testing before you believe them
- “More sensors is better.” More sensors is better only if they disagree usefully. A cheap extra camera that duplicates an existing angle changes nothing; an ultrasonic ring that covers the blind slice under the scan changes a lot.
- “LiDAR is the reliable option, cameras are the cheap option.” Both are mature. The cost difference now sits in compute and calibration rather than in the sensing principle — and the deciding factor is still your building, not the price list.
- “It works in any site, we will see on commissioning.” This is the claim to distrust most. A supplier who cannot name the conditions where their machine struggles has either not deployed much or is not saying.
Matching the sensor mix to your building
Before the technical conversation, classify your site — it narrows the options faster than any spec sheet:
- Glass, chrome or polished stone in the route. Expect to need visual sensing and, in the worst cases, floor markers. Test the exact atrium on a demo, not a showroom.
- Long, feature-poor halls or wide-open floors. Ask how mapping scales and whether markers or survey-based references are part of the plan.
- Dense racking, machinery or furniture. Geometry is rich, so laser sensing carries the position; the question shifts to whether the machine sees the low obstacles between the shelves.
- Crowds — retail, transport hubs, food service. Detection is not avoidance: what matters is whether the machine slows, goes around and announces itself rather than stopping to wait for a clear aisle.
- Multi-floor buildings. Navigation continues through the lift lobby, so ask how floors are handled — as one map or as independent work areas per level.
This is the same sequence as the buyer’s checklist: decide the job and the building first, then let the machine class and its sensor set follow. Choosing sensors first is how sites end up with a well-specified robot that cannot finish a lap of the atrium.
What to ask instead of “LiDAR or camera?”
- Which specific conditions make this machine lose its position, and what does it do when it does?
- What happens to the map when racking or displays move — and who re-maps, at what cost?
- In a crowd, does the machine slow and route around, or stop and wait?
- How does it detect glass, reflective floors and low obstacles — and can I test it on my own floor before ordering?
- What is the maintenance rhythm for sensors and lenses, and what does it cost per year?
A supplier with real deployments answers all five with site conditions rather than adjectives. That answer, not the word on the sensor list, is what your reliability actually depends on.
Where to go next
If you want the working detail — how sensing becomes mapping, localisation, planning and a cleaning pass — read the navigation technology explainer. If you are still choosing the machine class, the machine overview and the industry guide size the job first. And if you would rather test the question against your own building, send us the floor plan and the awkward spots — the glazed wall, the long hall, the crowded entrance — and we will tell you which configuration we would put on it and why.