How AI Sorting Works

Near-infrared sensors read the polymer

Most optical sorters identify plastics by how they reflect near-infrared light. Writing in Plastics Engineering, the Society of Plastics Engineers' magazine, in February 2026, Mariana Holguin describes the process: material on a fast belt is lit with near-infrared light between 800 and 2,500 nanometres, the reflected spectrum is matched to a polymer, and air jets blow selected items into a separate stream. The same article explains why black packaging escapes. Conventional carbon black pigment absorbs near-infrared light so strongly that the reflected signal is effectively zero, which leaves the sorter with no usable information. If you are comparing optical units such as the BHS NRT Optical Sorter, ask which materials and colours its sensors resolve.

Vision models add shape and context

Cameras paired with trained AI models recognise items by appearance rather than chemistry: a carton, a tub, a sleeved bottle or a battery that should never have reached the belt. That lets a system pick out categories a spectrometer struggles with, and it produces counts and composition data as material passes. Recognition is only as good as the training images, so a model built on one region's packaging may stumble on another's. Before signing, find out how new categories are taught to the model, how long that takes and whether the image data collected on your belt stays yours. Then run a trial on your own material at your normal belt loading, not on a cleaned demonstration sample.

Robot pickers trade speed for flexibility

Robotic pickers use suction cups or grippers on fast arms to lift single items off the belt, guided by vision systems. They can be retrained for new targets and fitted over existing conveyors, but each pick takes time and every arm has limited reach. A 2026 Waste Management paper by Haldankar and Schumacher, published through NIST, combined a literature review with industry interviews and found operators moving away from full robotic integration because of limits in speed, reliability and gripper effectiveness. If a robotic unit such as the AMP Robotics Cortex or BHS Max-AI AQC is on your shortlist, ask for picks per minute on material like yours, the share of attempts that succeed and how grippers cope with wet film and flattened containers.

AI as a standalone quality monitor

Detection can earn its keep without any arm attached. The same NIST paper reports a shift towards deploying AI systems as standalone detection tools, particularly for quality control and facility monitoring. Cameras over infeed, residue and product belts show what is arriving, what is being lost and how pure each output is, so supervisors can adjust belt speed, screen settings or staffing before a load is rejected. An analytics unit such as the Greyparrot Analyzer, listed from about $50,000 in October 2026, is built for this kind of monitoring. Start by deciding which decisions the data should change, place cameras where those decisions are made and agree how alerts reach the people on shift.

Sources and further reading

• Robotic Sorting and Artificial Intelligence in Material Recovery Facilities: A Mixed-Methods Study of Published Research and Industry Perspectives — National Institute of Standards and Technology (published in Waste Management).
  NIST-authored review and industry interviews on AI and robotic sorting in recycling plants, and why robots lag detection.
• When Color Becomes Waste — Plastics Engineering (Society of Plastics Engineers).
  How near-infrared sorters identify plastics and why carbon-black pigments leave black packaging invisible to them.

Common questions

How accurate are AI recycling robots?

Accuracy depends on the material mix, belt loading and how purity is measured, so headline figures from different makers rarely compare directly. Ask for results measured on your own material with the method stated, and check purity and recovery together, since one can rise as the other falls.

Do AI sorting systems replace manual pickers?

Usually they change the job rather than remove it. People still deal with items machines miss, clear jams, check quality and maintain equipment. Plan how roles will change, and involve pickers in trials, since they know which items cause trouble on your belts.