Robots Are Learning to Sort Clothes for Textile Recycling

The tag on a jacket says all cotton. The jacket disagrees. Somewhere in the lining there’s polyester, and nobody wrote that down. It’s the daily problem at textile sorting plants, and it’s why some of them let robots take the first look. 

Sorting used clothes by hand 

Say it comes in a truck with a few tonnes of other donations, tied up in a bag. Somebody has to decide what it’s worth. Wear it again? Turn it into something new? Or write it off as residual and send it to the lowest-value pile? 

Sympany, a Dutch textile collector, gets trucks like that every day. Sorting crews make the call well enough, but there’s a lot of cloth and very few seconds per item, and modern clothes don’t help. Labels fade. Blends are everywhere. A sweater that looks like plain cotton can hide polyester and elastane, and recycling gets messy fast when fibers are mixed up together. That’s why sorting sits at the front of the whole process, and why people want a machine looking at it. 

Scanning fabric with hyperspectral cameras 

The first stop is a scanner. Human eyes are incapable to differentiate between cotton and polyester, but a hyperspectral camera can. The smart camera can record how fabric soaks up infrared light, and an AI model can match that pattern to materials. At the University of Groningen, a team built a model through data from a specialized camera and used it to separate cotton, wool and polyester with utter accuracy. Their partner’s goal is a decision on every item in about a second: resell, upcycle or residual. 

Plant builders have their own versions. PICVISA says its machine pairs a near infrared camera with AI that reads composition, color and shape, then flicks each item into the right bin with jets of air. A system from Hong Kong lines up four AI modules along a conveyor, with robot arms checking type, fabric structure, material and color. Kingston University’s line takes about two seconds a garment. 

None of it is magic. A mix of fibers can fool a scan. Dark dyes soak up the very light the camera depends on, so a black coat is harder to read than a white tee. Some systems add a condition check for stains and tears on top. 

Handling garments and removing zippers 

After the scan, the jacket has to be moved, and cloth is a nightmare to grab. It folds over itself. It slumps across a gripper. It sticks out in odd directions. Groningen’s lab started with small robots that spread out towels, because a towel is about as easy as clothing gets. Jackets come later. 

Kingston is building humanoid robots that can fold and unfold garments, and its line gives cameras and robots another job: spot the zippers, buckles and other bits that aren’t fabric, and take them off. Metal and plastic don’t belong in fiber recycling, and they’re easy to miss in a heap of coats. For a jacket, the zip is probably the first thing to go. 

Student engineers have built a robot arm that sorts fabric by predicted type, whether wool, polyester or a blend. Their pitch is that automation cuts out repetitive hand work, which is where mistakes creep in. 

Training the AI with human input 

A model can’t tell a good second-hand coat from a tired one without seeing plenty of examples. In the Groningen project, the people who sort clothes today are the ones labeling images as suitable or unsuitable, so what they know about wear and quality slides into the model. It’s an odd job swap. Experts spend their days teaching the thing that will help them. 

Data is hungry work for the whole field. Systems need big, well-labeled sets to function, and clothes show up in endless styles, colors and states of wear. A model trained on one country’s donations might struggle with another’s. 

Kingston University’s sorting line 

One scanner and one gripper don’t make a plant. Kingston’s project joins the pieces into a single line. AI sorts garments by type, color and composition. Zips and buckles come off. UV light sterilizes the clothing, a machine pulls out long fibers that can be reused, and poor-quality textiles get removed before the rest head off for recycling. The team is aiming at up to 30 tonnes of clothing a week. 

Researchers at Fraunhofer in Germany opine that robotic sorting lifts throughput by as much as 60% compared to conveyor-based or semi-automated setups (which shows less energy and less hand work). They also developed digital twins and virtual copies of a sorting plant to test ideas on a screen before anyone cuts steel. Statistics like these are from labs and pilot lines; hence, a full-size plant may land somewhere different. 

Where sorted clothes go 

If the scan says cotton, the zip’s gone and the coat looks decent, it might go on a rail for resale. If it’s tired but clean, it heads for upcycling, or a recycling route that suits its fibers. If it’s a tangle of blended threads nobody can separate, it joins the residual pile, and engineers would love that pile to shrink. 

One idea is to let garments announce themselves. Digital tags such as RFID chips or QR codes could tell a robot what a piece of clothing is made of before it ever looks. That would turn a guess into a lookup. 

A jacket you’ve worn for six winters gets two seconds in front of a camera, and those two seconds decide whether it’s worn again, shredded for fiber or left behind. Not much time, for something that held up those winters.