NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

The Chinese Recycling Sorter That Tells PET From PVC — at 3 Tonnes an Hour

China's AI optical sorters run in plants from Japan to Italy. At 3 tonnes an hour they replace three workers a shift — yet they still cannot handle a bag.

2026-11-16 · 1145 words · NeuroAI
The Chinese Recycling Sorter That Tells PET From PVC — at 3 Tonnes an Hour

At a plant outside Tokyo, mixed household waste moves down a belt at 300 tonnes a day. Cameras and near-infrared sensors read each object as it passes, air nozzles fire, and a fraction of the stream is knocked into a different bin. The machine doing the reading was designed in China.

Waste sorting (垃圾分拣) is an unglamorous corner of robotics. It is also one of the few places where a Chinese machine has quietly become the default in developed markets that are fussy about who supplies their infrastructure.

What the machine actually does

The category is called an AI optical sorter (AI光选机). A belt feeds material under a bank of sensors — visible-light cameras, near-infrared spectrometers, sometimes laser profilometers — and a model classifies each item by material, colour and shape within milliseconds. A row of compressed-air valves then blows the target items off the belt at the right moment.

The physics has been around for two decades; the shift is in the classification. Older machines were rule-based and worked well on single-stream material — crushed PET bottles, say, all the same shape and colour. Household waste is not single-stream. A sorting line sees a shampoo bottle, a yoghurt cup, a printed film, a PET bottle with a full sleeve label, and a piece of PVC that looks exactly like PET under visible light but ruins a recycling batch if it gets through.

That last problem is why near-infrared matters. Different polymers absorb differently in the infrared, so a spectrometer can separate PET from PVC even when they look identical. What changed after 2020 is that the classification step moved from hand-tuned thresholds to trained models that can handle dirty, deformed, overlapping material — and that the whole sensor stack got cheap enough to put on a machine that sells to a municipal contractor rather than a national lab.

The numbers the company puts on it

Data Beyond Technology (弓叶科技), founded in 2018, is one of the most visible suppliers in this niche. Its own published figures, which have not been independently audited:

  • Thousands of AI optical sorters in service, across China, Japan, South Korea, Italy, Brazil, Mexico and parts of Southeast Asia.
  • Reference plants including a 300 tonne-per-day line in Japan, 200 tonnes per day in South Korea, 300 in India, and a 1,000 tonne-per-day installation in Italy.
  • Recognition accuracy improved from 93–95% in early models to about 99% today.
  • The smallest unit handles 3 tonnes per hour and, the company says, replaces three workers on a single shift or six across two shifts. The largest units run above 10 tonnes per hour and replace more than twenty.

Mo Zhuoya (莫卓亚), who leads Data Beyond, put the cost story in an interview with China Women's News in July 2026: the equipment went from a seven-figure price in RMB to a six-figure one — from roughly US$140,000 and up, down to something in the US$14,000 to US$140,000 band — as the company scaled and brought core components in-house. That is the sentence that explains the export record. European and Japanese sorting equipment has a long reputation and a service contract to match; a machine that costs a fifth as much and can be maintained by the plant's own technicians changes the maths for a mid-sized operator in Malaysia or Mexico.

Why this is a robotics story, not a recycling story

Three things make waste sorting an unusually good fit for machine vision, and they explain why it happened years before anything resembling a general-purpose robot arrived on the same floor.

The task is narrow. Every item passes the sensor bank in the same orientation, at the same speed, under the same lighting. You do not need a robot that understands a kitchen. You need a classifier that recognises forty material classes on a belt.

The economics are immediate. Sorting labour is repetitive, dirty, and hard to staff almost everywhere — the shortage is not a Chinese problem. A machine that runs three shifts without a lunch break pays back in months, not years, and the payback does not depend on a subsidy.

The output is measurable. A plant can weigh what came out of each bin and send it to a mill. Purity is a price. That gives the buyer a way to check the vendor's claim, which is rare in robotics sales.

There is a policy tailwind on the Chinese side too. The zero-waste city (无废城市) programme pushed municipal governments to build sorting capacity rather than just landfill, and that domestic base gave suppliers the volume to iterate — the same pattern that made Chinese solar and battery equipment cheap.

Honest limitations

The 99% accuracy figure needs a caveat that the marketing does not carry. Accuracy in sorting is measured per pass, on a defined material mix, and it is not the same as plant-level recovery. A machine that identifies 99% of what it sees correctly still produces a residue stream, and the residue is where the arguments with regulators happen.

The tonne-per-day figures are nameplate capacities from reference projects. I could not verify them against an independent audit, and plant throughput depends on input composition and on how many hours the line actually runs — few lines run at nameplate.

The worker-replacement numbers come from the company, in an interview. They are plausible arithmetic (one machine, one shift, three pickers) but they are not a labour study, and they ignore the technicians now needed to maintain the sensors.

I also cannot tell you how these machines perform on the two things that most often break a sorting line: black plastics, which absorb rather than reflect near-infrared light, and flexible film, which tangles in any mechanical handling. Both are known weak points across the industry, not specific to one vendor.

What readers can do now

  • If you run a materials recovery facility: pilot on your worst stream, not your best. Ask for a one-week trial on mixed household waste and weigh the residue yourself. A vendor demonstration on clean PET tells you nothing about your Tuesday shift.
  • If you are a city or procurement officer: write purity and residue targets into the contract, with penalties, rather than buying on throughput. Throughput is easy to hit and easy to game.
  • If you are a household: the machine's ceiling is set at your bin. A bottle with liquid in it, a bag with mixed waste inside it, or a greasy takeaway box will defeat a 99% classifier, because the classifier sees the bag, not the contents. Rinsing and not bagging is not a moral instruction; it is the input specification.

The interesting thing about this corner of robotics is that it never needed a humanoid. It needed a camera, a spectrometer, a fast valve and a business case — and it got all four about five years before the general-purpose robots arrived to collect the headlines.

Related coverage

Cite this article

NeuroAI (脑机网), "The Chinese Recycling Sorter That Tells PET From PVC — at 3 Tonnes an Hour," 16 November 2026, https://neuroai.site/a/na-robot-optical-sorter-export

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Humanoids & Robotics” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories