Technology
Different measurements.
One sorting decision.
A complex object is more than its surface colour. Matvision combines detection technologies to identify each piece before it reaches the sorting robot.
The limit of physical separation
Physical separators sort by one property.
Complex streams need more than one.
A shredder line already does a lot of the work. An air classifier removes the light fraction, an overband magnet pulls out the ferrous metals, an eddy current separator ejects the non-ferrous. Each of these machines separates on a single physical property: weight, magnetism, conductivity.
What comes out the other end is a stream where that logic no longer helps. Zorba is a mix of aluminium, copper, zinc, brass and lead of similar size and similar weight. E40 steel scrap is clean, except for the few pieces the magnet took along because they were attached to steel: electric motors with copper windings, aluminium profiles held by a steel screw, cables, rubber. Optical sorters take the next step, but they eject with compressed air and split the stream in two, one category per pass.
Two pieces that look the same.
A value ratio of one to twenty-five.
To a colour camera, a steel casing and an electric motor are two grey pieces of steel. X-ray transmission sees the copper winding inside the second one. That single measurement changes what the piece is worth, where it should go, and whether the lot it belongs to meets the buyer’s specification.
From sensing to sorting
Four steps.
One integrated system.
01
Sensing
Different sensors measure different properties of the same object.
02
Data fusion
Measurements are aligned to build a consistent picture of each piece.
03
AI decision
The combined data identifies the material and its sorting category.
04
Robotic sorting
Robots pick each object and place it in the appropriate output.
Matvision Refine · 3D walk-through
Inside the machine.
Follow the stream through the sensing block and the robot cells, the way each piece sees it.
Scroll to enter
01 · Infeed
Material on the belt
The stream arrives already largely clean. What follows is about the few pieces that should not be there.
02 · Matvision Sense
Read in one pass
The sensing block measures every piece as it goes by: relief, colour, and what the sensors see inside it. One pass, no second look.
- 3D PROFILOMETER
- COLOUR CAMERA
- LIGHT-CONTROLLED ENCLOSURE
03 · AI model
Three images, one prediction
The relief, the colour and the X-ray image of each piece are fused into a single input. The AI model reads them together and returns what the piece is, with a confidence score.
04 · Decision
The model gives each piece a class
At the exit of the sensing block, every piece carries a decision. Most stay as they are. A few are contaminants: those take a colour, and a destination.
05 · Delta robots
Only the marked pieces are picked
A delta robot drops its gripper and takes the piece that does not belong. Purification is not sorting: we remove the few rather than classify the many.
06 · Outputs
Each class to its own bin
The contaminants leave by their chute, one output per class. The cleaned stream carries on to the end of the belt, unchanged.
07 · The whole line
Read everything, pick a few.
Sorting in motion
See the sorting
in action.
Complementary sensors
Each sensor adds a different view.
| Sensor | What it contributes | Why it matters |
|---|---|---|
| Colour camera | Surface appearance and colour | Visual distinctions between materials. |
| 3D profilometer | Geometry, volume and thickness | Interpreting measurements and guiding the robotic pick. |
| X-ray transmission | Transmission through the piece and internal structure | Information hidden from a surface camera. |
| LIBS, when configured | Elemental composition | Finer distinctions between alloy grades. |
The choice of sensors depends on the stream and the separation required. Adding a sensor is justified by the value of the extra distinction it enables.
What makes it ours
Multi-sensor, multi-output.
The two words that describe the whole approach.
Matvision recovers valuable fractions from complex residual streams. We do not replace the shredder, the magnet, the eddy current separator or the optical sorter. We come after them, on the residual stream they cannot split further, and we define as many categories as that stream holds value.
Flexibility
The models are updated in operation and follow the composition of the stream as it changes. When the input stream changes, model performance is checked and the sorting settings are reviewed.
Complex streams
Purities that single-sensor machines do not reach, because they are decided on one property. That is where we work, and it is the only argument that matters for a plant that already has a sorting machine in the hall.
Depth of integration
Combining several sensors, a fusion layer, AI and robots into a system that holds an industrial pace takes years of development, test data and investment. That know-how is protected and it is the real barrier to entry.
What we do not promise
Precision is a dial, not a number.
If a line only sorts the pieces whose composition is certain, purity is excellent and a lot of valuable material passes through. If it recovers everything, purity drops. The right setting depends on what you sell and at what price, and it is set with you, on your material.
That is why we do not announce a theoretical return on investment. We announce a range, typically one to three years depending on the composition of the stream, and the method to confirm it: a performance study on your own material, with the KPIs we can commit to contractually. Throughput and purity figures on this site are orders of magnitude, always given with their condition, never commitments.
Performance is measured, not assumed.
Every stream behaves differently. Before we commit to purity, recovery or throughput, we validate the sensor configuration, AI model and operating conditions on your material. These results become the basis of the KPIs included in our final offer.
Spectra · line supervision and production analytics
What the line measures, you see.
A sorting line produces data at every step: what came in, what each sensor saw, what the model decided, what the robots picked. Spectra turns this into a live view and into production reports.
With the next machines
- Live status of the machine, sensors, robots and active model
- Pieces detected and pickable per second
- Throughput and composition of the stream in real time
- Production analytics by day, lot and material, in pieces, weight and percentage
- Traceability of the software version, model and thresholds behind every lot
- Operator and plant manager access, with Excel and CSV exports
- Machine events, stops and availability
On the roadmap
- Robot pick performance per piece
- Material analytics and a digital passport per lot, including size and weight distributions by family at line input
- Anomaly and drift detection
- Per-lot production reports
One platform, three products
The same sensing core,
three ways to use it.
Every Matvision machine is built on the same sensing and decision core, then sized to the job: characterise a stream, clean a stream, or split it into all its valuable fractions. Conveyors are part of the machine and part of the price, because our sensors need a controlled presentation of the material: speed, regularity, layer thickness, lighting.
Questions about sensors, robots or LIBS? Read the FAQ ↗