20 millimetres and a glass wall: what a cleaning robot's sensors are documented not to see

The Gausium Phantas is a compact cleaning machine for indoor floors, 585 mm long, 445 mm wide, 619 mm high, 65 kg, with four cleaning functions. The manufacturer's specification page, read in September 2026, lists exactly five entries under 'Sensing': '2D LiDAR, 3D Depth Cameras, RGB Camera, Anti-drop, Anti-collision'. No ultrasonic, no figure for the number of cameras, their range or the mounting height of the scanner. Under 'Movement' are the numbers that count on a site walk: gradeability 5 degrees, minimum pass width 650 mm, minimum pass height 650 mm, and the one line this article carries in its title: 'Min. Height of Detected Obstacles 20 mm'.

That line is valuable precisely because it is so rare. It says from what height an object on the floor is treated as an obstacle by the sensors. It does not say which of the four layers detects it, and it does not say what happens to an object 15 mm high: a flat cable, a tile edge, a document, a coat hanger. Below the documented threshold there is no assurance, and the machine may drive over it. Beneath it stands the sentence every manufacturer writes and operators rarely read: 'The specifications are based on the test results conducted by Gausium; actual performance data may vary in specific applications.' For comparison, the second catalogue type, the Pudu CC1: its manufacturer page names, for the Pro variant, 'LiDAR + Visual Fusion Positioning' as navigation and 70 cm as minimum path clearance, but no threshold for obstacle height. Where the value is missing it is not zero; it is unknown.

A 2D laser scanner sends a rotating laser beam out in one plane and measures the distance to the nearest object from the time of flight. It is the layer with which the machine builds its map and localises itself in it. Two limits follow from the design. The first is geometric: it sees only in its plane. What lies below or above the scan height, a table top at chest height or a pallet with overhang, does not exist for this layer. The second is optical, and for that there is an independent source.

Tibebu, Roche, De Silva and Kondoz of Loughborough University London published a measurement study on glass detection with LiDAR in the journal Sensors in 2021. Their starting finding is in the first paragraph: 'most modern environments contain glass, which is invisible to LiDAR'. The reason is physics, not software maturity: 'LiDAR is expensive and has major drawbacks when scanning in a transparent or specular reflective surface, such as glasses and mirrors. Hence, LiDAR sensors only account for diffuse objects.' The laser light passes through the pane and measures whatever stands behind it. A return from the pane itself exists in one case only: 'intensity peaks on a glass surface are only detectable where the LiDAR laser beams are perpendicular to the glass surface'. And the pane additionally produces false points: 'a significant amount of noise caused by virtual cloud points created from the reflection of objects nearby the transparent materials'. The authors collected their data with a Velodyne VLP-16 on a vehicle, not on a cleaning machine; but the properties described are those of the measuring principle, not of the vehicle, and apply to the scanner on the robot just the same.

For a building that means: a glass partition that reaches the floor is, at most points, absent from the scanner's map. The robot plans its route through it, and whether it stops is decided by the next layer.

Depth cameras supply what the scanner lacks: an image with height. They are the layer that detects a low obstacle in front of the machine, and the 20-millimetre figure most probably traces back to them, even if the datasheet does not say so. There are three common designs, active stereo, structured light and time of flight, and all three work with infrared light that has to come back from the scene.

Curto and Araujo of the University of Coimbra measured, in 2022 and likewise in Sensors, three cameras of these three designs under controlled light in front of glass and black surfaces. Two of their conclusions are directly usable for a cleaning machine. First: 'The transparency due to the aquarium glass walls do not affect the depth estimation in the three cameras. The wall is detected by all the cameras in a consistent way.' Two glass panes with air between them did not disturb the cameras because they did not see the panes; they measured the wall behind. That is the same finding as for the scanner, with different light: the glass door is not a surface for the depth camera either. Second: 'The black cardboard was a significant constraint for the SR305 and L515 cameras, being more significant for the SR305.' A black card on the wall took the structured-light camera and the time-of-flight camera to their limits; for the third, an active stereo camera, the study does not name this constraint, and it comes out best in the authors' overall assessment. Which design is built into a particular cleaning machine no datasheet says, and that is why the question belongs in the demonstration: not in the dealer's corridor but in front of the black plinth in your own lobby.

The RGB camera is the layer with which the machine classifies things, such as soiling or object types, and with which the manufacturer provides images to the operator. It delivers no distance and needs light; in a corridor where the lighting runs on motion sensors at night, it sees nothing between two triggers. The drop sensor, 'Anti-drop' in the datasheet, looks downward and is meant to stop the machine driving down a staircase or a loading ramp. The datasheet names neither the measuring principle, nor the step height at which it triggers, nor the floor colour it was tested on. The same goes for 'Anti-collision', usually a mechanical bumper: it is the last layer that engages when all the optical ones have failed, and it engages on contact, not before.

Four layers together make a machine that drives safely in an ordinary office building. The limits lie where two layers miss the same thing: a floor-to-ceiling glass wall is invisible to the scanner and transparent to the depth camera, and the bumper remains. A black floor covering in front of a staircase challenges the depth camera and the drop sensor at once, and for the second there is no documented assurance.

From the datasheet and the two studies a list can be derived that is worked through before the quote. It is deliberately short. First, glass to the floor: glass partitions, glass doors, shop windows, display cases, glass balustrades, each with the question whether a skirting or frame exists at scan height; where not, the spot goes into the map as an exclusion zone or gets a marker. Second, black and dark: black plinths, dark carpet edges, anthracite tiles before steps, matt-black furniture feet; here the demonstration decides, not the datasheet. Third, under 20 millimetres: cable protectors, doormats with a flat edge, door thresholds, expansion joint profiles, anything flat enough to lie under the documented threshold; the question is not whether the machine sees it but whether it may drive over it. Fourth, under 650 millimetres: passages between shelving, furniture and posts, lift doors, door leaves that do not open fully. Fifth, over 5 degrees: ramps, transitions between building sections, gradients towards drainage channels. Sixth, drop edges: stair descents, loading ramps, platforms, with the question to the manufacturer at what height and on what floor colour the drop sensor was tested.

Every item has one of three answers: exclusion zone in the map, structural change, or proof in the demonstration that the machine handles the spot. What has none of the three answers is a fault in the first month of operation, and that costs more than the walk-through.

The sensor entries and limit values come from the Gausium Phantas specification page and the Pudu CC1 Pro product page, both read in September 2026; manufacturers change these pages without notice, and a quote should fix the values in the contract rather than point to the website. The two studies are Sensors 2021, 21(7), 2263 (Tibebu et al., LiDAR-Based Glass Detection for Improved Occupancy Grid Mapping) and Sensors 2022, 22(19), 7378 (Curto and Araujo, An Experimental Assessment of Depth Estimation in Transparent and Translucent Scenes for Intel RealSense D415, SR305 and L515). Neither was collected on a cleaning machine; both describe the measuring principles built into such machines, not the software with which a particular manufacturer covers their gaps. Which depth camera design sits in the Phantas or the CC1 the manufacturers do not say; attributing the 20-millimetre figure to the depth camera is an inference of this article, not a manufacturer statement. And there is no measurement cited here of how these machines' drop sensors behave on dark floors; that is exactly why the question is on the list.