AI in Recycling: How Vision Systems Improve Waste Sorting
The verdict: the robots can see the rubbish. Whether that saves the planet is a different question.
NVIDIA's November 10 report showcases four startups using computer vision to sort waste — plastics, household recyclables, used clothing. The numbers are eye-catching: 95% recognition accuracy, 40 billion objects identified, a trillion bounding-box detections. But read the fine print and the story gets more interesting — and more honest — than the press-release version.
One caveat up front: all four companies are in NVIDIA's Inception startup program, and the report comes from their computing supplier. These are company estimates, pilot measurements, and future targets — not an independent comparison. Keep that in mind while the numbers fly.
95% accuracy is not 95% recycled
Taiwan-based AIRECO runs vision AI on a pilot line in Tainan, hitting above 95% recognition accuracy on plastics. That's a real result — under test conditions. What the report doesn't say is what share of all incoming material actually gets recovered, or what the plant economics look like. Recognition and recovery are related tasks, not the same measurement.
AIRECO's ambitions are bigger: a full production line planned for 2027, capacity up from 18,000 to 48,000 tons a year, operating costs down 85%. Targets, not results. Judge them when the line runs.
The conveyor belt doesn't wait for your model
EverestLabs AI puts the timing problem front and centre. Its ~150 robots across North America and Australia have to identify an object and pick it before it rolls past on a moving belt. A correct classification that arrives a second late is worth nothing.
The company estimates its systems have identified about 40 billion objects, recycled roughly 25,000 tons of material, and avoided about 30,000 metric tons of CO₂. Cumulative company estimates, methodology undisclosed — impressive if true, unverifiable from this report.
Sometimes the win is just seeing clearly
Greyparrot doesn't pick anything up. Its Analyzer watches the belt from above and tells operators what's actually flowing through their plant — material and brand, at up to 4 metres a second, in dust and poor light. Manual sampling covers less than 1% of the stream; Greyparrot claims over 300 systems in 20+ countries and more than a trillion detections.
The headline stat: checking 15 tons of waste drops from 375 hours to six. Facilities report 10–30% efficiency gains; Amcor reportedly lifted food-grade plastic yield by 12%. Attributed outcomes, not a universal promise — but the measurement angle is the most defensible use case in the whole report. You can't improve what you can't see, and until now operators were mostly guessing.
A torn seam changes everything
Berlin's reverse.fashion is the subtlest case. Sorting used clothing isn't about material type — it's about value, and value lives in small details: a defect, a construction quirk, a style. Each garment passes through more than a dozen vision models at 3 metres a second, and gets assigned to its most valuable category. The CTO estimates it sorts in an hour what used to take a week — his estimate, with the assumptions unstated.
The takeaway: don't confuse seeing with solving
The honest contribution here is legibility: waste streams you can actually measure, and in some cases sorting equipment that acts on what it sees. Plastics separated finely enough for closed-loop recycling. Garments routed to their best use. Operators finally looking at the whole stream instead of a 1% sample.
But the sustainability case depends on what follows recognition — material actually recovered, feedstock actually reusable, garments actually going somewhere valuable. The report gives concrete examples and attributed results along that path, and it doesn't pretend AI alone guarantees circularity. Neither should anyone else.
The practical test for any operator: match the evaluation to the outcome. A plastic sorter needs recovery evidence, not accuracy percentages. A monitoring system needs to prove its data changes decisions. A garment classifier needs to prove useful categorisation, not just speed. And nobody should be selling a general ROI promise for AI in recycling until costs, maintenance, payback periods, and model errors are on the table. They're not — yet.
Go deeper
Read the full analysis on Agentic Affairs: AI in Recycling: How Vision Systems Improve Waste Sorting
Originally published on Agentic Affairs (https://agentic-affairs.com)
