Sensors, robots and LIBS, in plain terms.

Sensor-based sorting becomes particularly relevant when conventional mechanical separation has reached its limits.

Mechanical processes can efficiently separate materials based on properties such as size, shape, density or magnetic behaviour. But when different valuable materials share similar physical properties, these methods may no longer provide the selectivity required to produce high-value fractions.

Sensor-based sorting adds another layer of information by identifying individual objects based on their visual, physical or chemical characteristics. It is particularly relevant when:

• mechanical separation cannot achieve the required purity;
• materials with similar physical properties need to be distinguished;
• manual picking needs to be automated;
• labour availability or working conditions make hand sorting difficult to sustain;
• several different valuable fractions need to be recovered;
• plant footprint is limited and multiple successive separation stages would be impractical.

The objective is not to replace efficient mechanical separation, but to add sensor-based sorting where additional information creates additional value.

Read: Targeting the hidden part of the waste iceberg ↗

A shredder line already separates a lot: the magnet takes the ferrous metals, the eddy current separator ejects the non-ferrous, the air classifier removes the light fraction. Each of these machines separates on a single physical property. What comes out at the end is a stream where that logic no longer helps, and only a measurement of what each piece is made of can split it further.

A colour camera reads the light that comes back off a surface. It cannot tell how thick a piece is, what sits under the paint or the oxide, or which elements it is made of. Two metals can look alike under industrial lighting, and a steel casing and an electric motor are two grey pieces of steel until X-ray transmission sees the copper winding inside.

They are not competing technologies.

The recycling industry often refers to AI-based sorting when conventional RGB cameras are combined with computer vision and deep-learning algorithms, while sensor-based sorting usually refers to technologies such as X-ray, infrared, hyperspectral, LIBS or XRF that measure specific physical or chemical properties of the material.

In practice, this distinction is increasingly artificial. At Matvision, AI is used to interpret data from multiple types of sensors. Instead of applying deep learning only to colour images, our models can exploit information about an object’s appearance, geometry, density, spectral response or elemental composition.

This combination allows the AI to learn from richer information and enables the most appropriate sensing technology to be selected for each sorting problem. The question is therefore not AI or sensors, but rather: which sensors provide the useful information, and how can AI extract the most value from it?

Read: AI-driven smart sorting, the evolution beyond traditional sensor-based sorting ↗

Because high-resolution material identification only creates its full value if the sorting system can act on that information.

Different sensors provide different information. RGB imaging describes appearance, 3D sensing provides geometry, X-ray technologies provide information related to density or elemental composition, and LIBS can identify individual elements and alloying elements. Combining these signals can therefore enable much finer material classification than relying on a single measurement.

But recognition is only half of the problem. A LIBS system, for example, can distinguish numerous alloy families based on their elemental composition. If the downstream separation system is designed to produce only a small number of output fractions, much of that classification capability remains unused.

This is where multi-output robotic sorting becomes particularly powerful. Instead of reducing a rich classification result to a binary decision, individual pieces can be directed towards multiple different material fractions in a single sorting stage.

Matvision therefore combines multi-sensor recognition with multi-output sorting, so that increased intelligence on the sensing side can translate directly into increased material recovery and value on the sorting side.

See Matvision Grade, one output per category ↗

Yes, but its relevance strongly depends on the material and the sorting objective.

Hyperspectral imaging is particularly powerful for materials such as plastics. Different polymers can exhibit characteristic absorption bands associated with specific molecular bonds. Their spectral response can therefore provide a real chemical fingerprint and make it possible to distinguish materials that may look almost identical to a conventional colour camera.

Metals behave differently. Their optical spectra generally do not contain the same type of distinctive molecular absorption features, so hyperspectral imaging does not automatically provide direct information about alloy composition.

This does not make spectral imaging irrelevant for metals. Surface condition, oxidation, coatings and other optical characteristics can still provide valuable information for certain sorting tasks. In some applications, multispectral imaging can therefore be particularly interesting: it captures selected spectral bands beyond conventional RGB while maintaining a simpler and more industrially practical architecture than a full hyperspectral system.

When direct elemental or density information is required, technologies such as LIBS, XRF or XRT may provide more discriminating information.

At Matvision, we therefore do not start from a preferred sensor technology. We start from the separation problem and select the sensing combination that actually provides useful information about the material.

Because the number of outputs should not limit the classification. An air-jet sorter separates one target from the flow, or a few fractions within its mechanical design. A robot places each identified piece in the output where it is worth the most, which is what lets several grades leave a mixed stream in a single pass.

No. Both can work on the same line: robots on the pieces that need placing, air on the ones that only need removing, or on those that are small, fast or too abundant to be worth an arm. The split is decided by the performance study on your material.

It follows the stream, not a catalogue. Matvision Refine uses one to four robots, sized to the contaminant to take out; Matvision Grade uses five to twelve, one output per category worth splitting. A line can start with fewer arms and grow with the stream.

Laser-induced breakdown spectroscopy. A focused laser pulse vaporises a tiny amount of material, the vapour becomes a plasma, and a spectrometer splits the light that plasma emits into lines that name the elements present at the point of impact.

Because it reaches light elements such as magnesium, silicon, aluminium, lithium and carbon, and that is often where the value of a fraction sits: 5xxx against 6xxx aluminium is a question of magnesium and silicon, cast against wrought aluminium a question of composition, 304 against 316 stainless a question of alloying elements. The first pulses also remove the coating or oxide, so the later ones read the material underneath.

Neither, because they do not answer the same question. XRF is strong on heavier alloying elements such as copper, zinc, nickel or chromium, but its signal is the sum of everything it crosses, coating included. LIBS removes the surface and reaches the light elements, but it describes a crater of a few hundred micrometres and needs more calibration effort to stay reproducible. Their blind spots are opposite, which is why the choice depends on the separation.

Where the camera tells it to. Imaging locates each piece, its edges and its visually distinct regions, and the shots are placed on those regions rather than on a fixed pattern. A single crater on the copper insert of a steel body would be confidently wrong about the piece; shots chosen by the camera turn a local measurement into evidence about the object.

Yes, and that is the engineering problem to solve. A confidence score is not a probability: a classifier always ranks the classes it knows, even for a piece it has never seen. Calibration, meaning how far stated confidence matches observed accuracy, can be measured and corrected, and a system that knows when not to answer is worth more than one that always does.

Read: AI in recycling, constraints and challenges ↗

We measure it before anything is written into an offer. A performance study starts with two big bags of your material: we measure its composition piece by piece, set the purity we can hold against the material you would lose with the contaminants, and write both into the offer as the figures we commit to.

Yes. The six families on the use cases page are the ones we meet most often, not the limit of what the machines can do. A stream outside them becomes a project: we measure your material, define the sensor combination and the outputs it justifies, and build the line around that.

The method does not change: two big bags of material, a performance study, figures measured on your own stream, then a configuration that makes economic sense. What changes is that the result is specific to you rather than taken from a catalogue.

See what that looks like ↗

Yes, and that is part of how the technology keeps moving. Matvision came out of the GeMMe laboratory at the University of Liege, takes part in European collaborative R&D projects, and works with research teams and industrial groups on separations that have not been demonstrated yet.

In practice that can be a Matvision Sense unit installed on a pilot line or in a laboratory to collect real data, a characterisation campaign on material you send us, or a full line built for a single process. Consortium partners looking for the sensing and sorting part of a project are welcome to write.

Write to the team ↗

The price of a sorting machine depends directly on the configuration required for the application. Key cost drivers include the type and number of sensors, the number of robots, the required throughput, and whether the project requires a dedicated conveyor or other specific mechanical integration.

For this reason, we do not start with a machine configuration and try to make it fit the application. We first evaluate the material, the sorting objective and the expected value creation, and then define the configuration that makes economic sense.

Before recommending an investment, we assess the business case and expected return on investment. The applications presented on our website correspond to sorting problems for which a viable economic case has been identified. The return on investment can come from several sources:

• Labour automation: reducing dependence on manual picking and the associated labour costs.
• Consistent product quality: maintaining stable sorting performance and reducing the risk of quality penalties or downgraded material, particularly when staffing is difficult.
• Higher production capacity: enabling longer operating hours and reducing the limitations associated with manual sorting.
• Higher material value: separating materials according to criteria that may be difficult or impossible for a human operator to assess consistently, producing cleaner and more concentrated fractions with a higher market value.
• Improved recovery: identifying and recovering valuable materials that would otherwise remain in a lower-value fraction.

Ultimately, the relevant question is not only “How much does the machine cost?”, but “How much additional value can the sorting operation generate?” This is why Matvision evaluates the business case before confirming the final machine configuration.

Not sure whether sensor-based sorting makes economic sense for your material or process? Contact our team to review your application with our sorting experts and assess the technical feasibility, expected value creation and potential return on investment.

Talk to the team about your application ↗

Matvision brings together robotics, artificial intelligence, electronics, mechanical engineering and materials science, in a team small enough that everyone sees what the machine does with their work. Passion, creativity and a taste for problems nobody has solved yet count as much as the line on a diploma, and we would rather the work stayed enjoyable.

See our current openings ↗

If none of them matches you and you would like to put your skills at the service of a cause worth the effort, write to us with your CV. A good profile that arrives before the position exists is how most of this team was hired.

Send us your CV ↗