Stereo machine vision has a structural problem: two cameras cannot occupy the same point in space. The classical solution is a beamsplitter cube that shares one optical path between two sensors — and as 3D profiling, binocular picking and structured-light systems have moved into mainstream industrial deployment, the cube has quietly become a volume component.

It is also a component where several specification details decide whether the system works, and they are the details most RFQs leave out.
Why cubes and not plates
In a stereo rig, the two cameras must see identical geometry. A plate beamsplitter displaces and astigmatically distorts the transmitted beam in anything but a perfectly collimated path; a lens system is never perfectly collimated everywhere. A cube — two right-angle prisms cemented with the coating at the diagonal — has a glass path only, no second-surface ghost, and symmetrical behaviour for both outputs. For precision 3D geometry, that symmetry is the reason the cube wins despite its higher price.
Detail one: the cement line
The cube's diagonal is a cemented interface, and it is the component's most common failure point. Cement outgassing under heat, delamination under thermal cycling, and yellowing under strong illumination over years — these show up as transmission loss and veiling scatter that no incoming inspection catches at room temperature. For systems that live on a production line with warm lighting and temperature swings, buyers should ask for the cement type, its temperature limits, and thermal cycling qualification data. Optical-contact (cementless) bonding exists for demanding cases and removes the failure mode entirely at a cost premium.
Detail two: the prism, not just the coating
A cube is a thick glass block in the optical path. Its material homogeneity and the polish of all four outer faces contribute wavefront error that the coating datasheet says nothing about. In a 3D measurement system, wavefront error through one camera's path and not the other's produces a systematic depth bias. Transmitted wavefront, specified per output path, is the specification that keeps the two cameras geometrically identical.
Detail three: ratio where the sensor actually sits
The cube's split ratio varies with wavelength, angle and polarisation — as covered generally in beamsplitter selection — but in a stereo rig the practical consequence is exposure mismatch between the two cameras. If each camera compensates independently, the 3D reconstruction inherits a systematic brightness difference that uncalibrated algorithms turn into depth artifacts. Requesting a tight ratio tolerance at the sensor's illumination band, matched across both outputs, is cheaper than building the correction into software.
Detail four: thermal drift
The two prisms are usually the same material, which keeps thermal expansion matched. But the coating's ratio drifts with temperature, and a cube specified at 50:50 at 23 °C may sit at 52:48 inside a 45 °C housing. For systems with a warm enclosure and long duty cycles, the temperature range belongs on the drawing with the expected drift stated.
What belongs on the cube drawing
- Ratio and tolerance at the named wavelength band and angle, both outputs.
- Polarisation behaviour, if either sensor or illumination is polarised.
- Transmitted wavefront per output path, and surface flatness of the outer faces.
- Cement type, temperature range and cycling qualification — or optical-contact specification.
- Prism material grade and any striae requirement.
- Physical dimensions with the cube's glass path noted for back-focus calculation.
Beamsplitter cubes are bought as catalogue items and specified like catalogue items — a part number and a ratio. The systems that fail on the production floor are usually the ones where the cement, the prism or the thermal range turned out to matter. A cube for a stereo camera is not a commodity; it is the geometric anchor of the whole instrument.
Product page: beam-splitter cubes and plates — specifications, coating options and lead times.