Sorting used textiles with AI, NIR, and a recycling framework

In a circular textile value chain, one decision governs everything downstream: can this garment be reused, or does it go to recycling? Today that call is made mostly by hand. A person looks at a garment, judges its condition, guesses at its fibre composition, and sorts it. That works at current volumes. It does not work at the volumes coming, and it does not produce the data a circular economy needs in order to function.

We worked with Wargön Innovation on an automated alternative, as part of the project Ramverk för cirkulära textilier (Framework for Circular Textiles), funded by Vinnova. Our role was to analyse and develop an automated sorting process, and to work through how a platform for the collected data could be built. This note is what we learned.

Why one sensor is never enough

The instinct is to reach for a camera. Image recognition is good at what a person sees: garment type, colour, visible wear, stains, missing buttons, a broken zip. It is completely blind to what a person cannot see, which is what the garment is actually made of. A cotton shirt and a cotton polyester blend can be visually identical and belong in entirely different recycling streams.

Near infrared spectroscopy (NIR) answers the composition question. Different fibres absorb and reflect infrared light differently, so a spectral reading gives you a material signature. NIR in turn is blind to everything the camera is good at. It cannot tell you the garment is torn, or that it is a jacket rather than a bedsheet, or that the colour makes it commercially undesirable.

So the system reads a garment through several instruments at once: image recognition for type, colour, and condition, NIR for material composition, and weight. Each measurement is weak alone. The useful object is the combined record.

Where it gets difficult

Combining sensors sounds like plumbing. It is not, because the sensors disagree.

• Blends are the hard case. A pure fibre gives a clean spectral signature. A blend gives a mixture, and the proportions matter enormously for whether a recycler will accept it. Confidence has to travel with the reading, because a low confidence composition estimate is a different business decision from a high confidence one.
• Contamination and coatings distort readings. Prints, coatings, moisture, and dirt all sit between the sensor and the fibre. The reading is of the surface, not the intent.
• The instruments do not agree on what an item is. Reconciling one physical garment across several instruments, each with its own timing and failure modes, is most of the engineering.
• Ground truth is expensive. Knowing what a garment really contains means lab testing. That constrains how much labelled data exists to validate against, which shapes what the system can honestly claim.

This is why the project used real garment data. GinaTricot contributed data for the garments so the process could be simulated against genuine items rather than a clean synthetic set. Synthetic data would have made the results look considerably better and taught us considerably less.

Sorting toward demand, not just toward material

The part of this project we found most interesting is the part that is not really a sensing problem at all.

A garment is not recyclable in the abstract. It is recyclable if somebody will take that specific material, in that condition, in that volume, at a price that works. A sorting system that only classifies material composition produces neat categories that may have no buyer. That is how well intentioned recycling programmes end up with warehouses of correctly sorted material nobody wants.

So we built a knowledge base of potential buyers and what each of them actually accepts, and made it part of the sorting logic. The system does not only ask what this garment is made of. It asks who wants this, which turns sorting from a classification task into a matching task between supply and demand.

A material is only recyclable if someone is buying it.

Those readings then match against the recyclability and circularity framework developed in the project, so a garment's measured attributes connect to an assessment the whole value chain has agreed on rather than to one company's internal rules. This is also where Digital Product Passport (DPP) becomes practical rather than theoretical: once colour, weight, and material composition are captured as structured data at the point of sorting, there is something real to carry forward.

Why the data platform mattered as much as the sorting

Every garment through the line produces a record: what the instruments read, what the system concluded, how confident it was, and where the item was routed. That record is the asset. It is what lets you audit a sorting decision, prove circularity claims, and improve the models with real outcomes rather than assumptions.

It is also why we care where it runs. Sorting data describes commercial relationships, supplier volumes, and material flows, which is competitively sensitive information across an entire value chain. Our position on this is the same as it is for every client, and it is set out in our reference architecture for sovereign ML: the data stays with the organisation that generated it, and the models run where the data is.

Where this stands

Findings from the project were presented at Science Park Borås, at DO-tank Center, jointly by Gabriella Engström of Wargön Innovation and Andrija Ilić, co-founder of Quince AI, covering the possibilities and the lessons from connecting different technical instruments to identify data about a textile.

What this project produced is an analysed and developed sorting process, validated in simulation against real garment data, with the reuse versus recycle routing decision as the target. That routing call is the coarsest decision in the chain and the one most amenable to automation, which is exactly why it was the right thing to attempt first.

What it has not yet produced is a sorting line running in production with numbers we can stand behind. There is no accuracy figure or throughput rate in this note, because a figure only means something alongside the sample size, the ground truth method, and the confusion matrix behind it. Reporting accuracy on a routing decision without saying how the hard cases were counted is how sorting projects end up promising more than the line delivers.

We mention this because the temptation with an innovation project is to publish the demo as though it were a deployment. The rules we publish by say a project stays a work note until it carries both a name and a verified number. This one carries the name.

What Wargön Innovation said

What we would carry into the next one

Three things from this project apply well beyond textiles, and we have used all three since.

• Fuse sensors, and keep the confidence. Any physical sorting problem, whether textiles, plastics, or construction waste, has the same shape: several partial instruments and no single one that answers the question. Carry uncertainty through to the decision instead of flattening it to a label.
• Model the downstream market, not just the material. The classification is only worth what the offtake is worth.
• Treat the data record as the product. The sorting decision is consumed once. The structured record of why it was made keeps paying, in audits, in circularity reporting, and in the next model.

If you are working on automated sorting, materials classification, or circularity reporting for Swedish industry or the public sector, the ML Feasibility Sprint is two weeks at a fixed price, and it ends with an honest answer about whether the machine learning is worth building at all.

Ramverk för cirkulära textilier is funded by Vinnova. Project partners included Wargön Innovation, Science Park Borås, and GinaTricot.