How Cleaning Robots Navigate: LiDAR, SLAM and the Limits of the Sensors

Everything a cleaning robot does with its sensors falls into three jobs that are easy to confuse. First, mapping: once, at commissioning, the robot records the geometry of your building. Second, localization: during every run, every second, it has to answer the question "where exactly am I on that map" — this is what decides whether the coverage pattern stays clean or drifts into overlaps and gaps. Third, obstacle handling: reacting in real time to whatever is in the way today and is not on the map at all.

SLAM — simultaneous localization and mapping — is the name for doing the first two at the same time. The plainest official definition we found is in Kemaro's own factsheet for the K900 Gen II: the robot "simultaneously creates a map of its environment and estimate its spatial position within this map". Pudu names the same thing as a stack: the official MT1 brochure gives the navigation method as "VSLAM+Marker+Lidar SLAM", the MT1 Max page as "VSLAM + Marker + 3D LiDAR SLAM". The "marker" in the middle is literal: physical positioning markers placed in the environment, which the visual navigation uses as fixed reference points — a component that almost always drops out of comparison tables.

Vocabulary differs more than the technology does. The official Nexaro datasheet for the NR 1700 calls its mapping technology "AI-Mapping" and its main sensor a long-range laser distance sensor (LR-LDS); the word SLAM does not appear on the datasheet at all. Nexaro's own FAQ, meanwhile, states that "thanks to smart SLAM algorithms and sensor data fusion, the robot can also navigate in larger areas". Same family of methods, two vocabularies — which is why SLAM shows up in some comparison tables and not in others, without anything technical being different.

And not every machine plans an area for itself. SoftBank states that the Whiz "vacuums along programmed routes and responds to obstacles and other changes to the environment": routes are set up for it, and the autonomy sits in following them safely rather than in planning coverage. Both models work. They just fail differently — and they need different things from you at setup.

A 2D LiDAR measures distances in one horizontal plane at a fixed height. It is precise, fast, robust in the dark, and structurally blind to everything above and below that plane: the low pallet corner, the cable drum, the shelf edge protruding at chest height. Manufacturers know this and say so. Kemaro describes the K900 Gen II vision system as "a vigilant 2D LiDAR 'crown'" measuring distances across 270 degrees, "and 3D camera fills in the lower blind spots" — the second half of that sentence is the whole argument for 3D sensing, written by a manufacturer selling a 2D scanner. The assessment rests on Kemaro's own documentation.

That is what 3D LiDAR and depth cameras add: volume instead of a slice. For the MT1 Max, Pudu's launch release states that "equipped with 3D LiDAR and multi-sensor fusion, the MT1 Max can 'understand' complex environments with centimeter-level accuracy and detect objects up to 150 meters away". The CenoBots S5, an industrial sweeper available through us but newly onboarded, is specified with a 32-beam 3D LiDAR at up to 150 m range on a 100 TOPS NVIDIA platform, explicitly to detect and avoid forklifts and AGVs. And the CC1 shows what mid-size multi-sensor fusion looks like in practice: its official brochure diagram lists a main LiDAR, a solid-state LiDAR, two RGBD cameras, a top-view camera, an ultrasonic radar, a cliff sensor and a bumper sensor.

Cameras buy perception and cost you a data conversation. Two manufacturers make the opposite choice deliberately. Nexaro's NR 1700 carries an LR-LDS, a bumper sensor, eight drop sensors, two magnetic field sensors, a wall-following sensor, incremental sensors and an IMU — and no camera. Adlatus writes in its SR1300 documentation that it is one of the few manufacturers to forgo high-resolution cameras in navigation for data protection reasons, using 2D and 3D lidar sensors instead, and that environmental data is recorded only as coordinates; the machine also runs fully self-contained without a WLAN connection or continuous internet. What that trade-off means for works councils and DPOs is a topic of its own — see data protection with cleaning robots.

Underneath all of that sits an unglamorous close-range layer that decides most real incidents: ultrasonic sensors for surfaces the laser struggles with, cliff sensors at edges and steps (Nexaro has eight; SoftBank states that Whiz safety sensors "detect and maneuver around people, objects, and cliffs"; Kemaro's FAQ answers for its robots generally, not for the K900 specifically, that drop prevention is "mechanical, significantly safer than the optical one"), and the bumper as the last instance. Pudu puts the honest sentence in its own CC1 operating guide: "Although the robot features automatic obstacle avoidance, there may be blind spots."

Glass and mirrors are the classic failure, and the manuals are unusually candid about it. Pudu's CC1 Operation Guide V4.2 warns that "pure black surfaces (such as baseboards), mirrored surfaces (such as walls), or fully transparent surroundings with a height between 14-22cm" may interfere with the robot's lidar reflection and cause abnormal movement, and recommends having the situation evaluated and, if necessary, modifying the environment with reflective stickers or materials. Read that height band again: 14–22 cm is where the laser plane sits. What stands at exactly that height in your building decides more about navigation than the marketing headline does.

Gausium goes further in the Phantas manual: if the environment is surrounded by glass walls or other highly permeable materials, "some of the S1 PRO sensors will fail to work", and using the robot there is not recommended. The troubleshooting table lists glass and transparent materials as a cause of a map that is "not clear or is ghosted" and gives the documented workaround — draw virtual walls to bound the map. This is a documentation-based assessment; but a manufacturer that writes its own sensor limit into the manual deserves the credit, because the alternative is a buyer discovering it on day three.

Our own machine is documented on the same question, and its answer comes in two layers. Nexaro's FAQ states that the laser distance sensor of the NR 1700 "may not detect some glass doors or glass walls", that the robot "detects any such transparent objects with its bumper sensor and navigates accordingly", and that glass can additionally be marked in the map as a no-go zone. The manual adds the inverse case: glass floors can be read as a drop, which stops the machine. So camera-free navigation does not exempt a robot from the glass problem — it moves the problem from the laser to the bumper, and the bumper solves it by touching the glass.

The second limit is the opposite problem: too little structure rather than too much. Laser localization needs geometry to match against. Nexaro's datasheet therefore specifies a maximum featureless open area of 12 m × 12 m for the NR 1700, alongside a maximum cleaning area of 1,000 m² per job, and the FAQ states the laser distance sensor detects surfaces up to 3 metres away. On the reach from the dock, the official sources disagree: the February 2025 datasheet footnote says a maximum distance to the charging station of 35 m, while the FAQ says the robot reaches points within a radius of up to 32 metres. Plan with the smaller number. In practice this is why an empty warehouse hall can be harder for a robot than a cluttered office.

The third limit is mundane and causes the most aborted runs: what lies on the floor. Pudu's CC1 guide asks for cables to be put away in advance so the robot does not drag them, specifies a minimum travel width of 70 cm with 75 cm preferred for long passages, and tolerates protrusions up to 8 mm while cleaning or 20 mm when only passing through. The MT1 brochure gives a minimum path clearance of 75 cm and names 20 mm thresholds and 35 mm grooves. The Phantas needs a 650 mm passage, and its 20 mm figure is easy to read backwards: Gausium's spec row is "Minimum Height of Detected Obstacles", the smallest thing the machine can see rather than the largest it can climb, while the manual separately permits doorsill protrusions of up to 3 cm. Its slope figure is a conflict between two official documents — the spec page says 5°, the manual forbids use above 8°. Chair legs, cable drums, pallet corners and a too-narrow aisle at the end of the rack: that is the list deployments actually fail on. Finally, sensors have to stay clean and dry — Gausium's troubleshooting names a dusty or blocked laser sensor as a cause of poor maps and prescribes a lint-free cloth, and states that Phantas sensors do not work properly outdoors in rain, fog or snow. Pudu, by contrast, claims in its official MT1 Max product deck stable mapping and positioning in heavy dust, fog and sandstorms, and recognition of scenarios from ultra-high ceilings and glass roofs to semi-open areas. Note precisely what that is: a claim about mapping stability, not a promise that glass surfaces are detected as obstacles. Pudu makes no such promise, and we do not either.

The demo most suppliers offer is a clean run through an empty hall at seven in the morning. That proves the drive works. It proves nothing about navigation, because navigation is only tested by disorder. Ask for the run in your worst hour, in the aisle you would rather not show. Before that, measure the narrowest passage you expect the machine to use and compare it against the manufacturer's stated minimum — 70 cm for the CC1, 75 cm for the MT1, 650 mm for the Phantas. Then put the real obstacles down: the cable drum, the pallet with the torn film, the stacked chairs, the cleaning trolley someone parks in the corridor every night.

The decisive test is not the first lap but the third. Let the machine run the same area three times, and change something in between — move a pallet, open a gate that was closed during mapping, park a trolley across the route. What you are watching for is whether coverage stays the same: whether the robot re-localizes on the changed map and cleans the same square metres, or quietly starts leaving strips out. Then ask for the evidence in writing. The fleet platforms document runs: PUDU Link delivers visualized cleaning reports and notifications when the bin is full, Kemaro's web app and Sphere cloud produce digital cleaning reports with interactive maps, the Nexaro HUB shows live mapping and a digital cleaning record, and Adlatus emails a fully automatic protocol after every run that is meant to hold up for billing and audits. A supplier who cannot show you three comparable coverage maps from three runs is asking you to buy on faith.

In halls with traffic, visibility is part of navigation, not a decoration. Kemaro added the BlueSpot in the Gen II specifically to alert employees and forklift drivers to the robot's movements with a blue light, including in areas with limited visibility or beneath conveyor systems; the CenoBots S5 is specified to identify forklifts and AGVs and plan avoidance paths; Pudu gives the MT1 Max a 1.2 m beacon for pedestrian visibility. Which of those approaches actually fits mixed traffic in a warehouse is the subject of our sweeper robot comparison. In offices, the same test looks different: glass partitions, chair legs and cable ducts under desks, plus staff who move furniture without telling anyone.

More sensing earns its keep where the environment changes while the robot works. Mixed traffic, high ceilings, semi-open areas, seasonal layout changes: that is the case for 3D perception, and it is why the MT1 Max carries a 3D LiDAR on top of the MT1's navigation stack. In a warehouse where pallets move hourly and forklifts share the aisle, the extra sensing is not a luxury — it is the difference between a run that completes and a robot that waits for someone.

Where the map is stable and the space is furnished rather than dynamic, the simpler machine often wins. The NR 1700 navigates on a laser distance sensor with no camera at all and cleans furnished office floors that way — quiet, camera-free and defensible in front of a works council. Its limits are equally clear and equally documented: 12 m × 12 m of featureless open space, 1,000 m² per job. Between the two sits the mid-size all-rounder: the CC1 with its multi-sensor fusion in retail and mixed-floor environments where furniture is dense but the building does not reshape itself weekly.

The honest closing frame: sensors decide whether the run finishes unattended. The cleaning head decides whether the floor is clean. Nobody has ever been happy with a robot that navigates brilliantly and scrubs badly, and the sensor list on a datasheet tells you nothing about the second half. Choose the machine class for your floors and your area first — our guide on which cleaning robot fits which operation walks that decision — and use the navigation questions in this article to check whether the machine you chose will actually survive your building.