Sorting recyclable material is a harder problem than it appears, and it is worth understanding why, because the difficulty is instructive. A conveyor carries a continuous stream of mixed material past at speed. The objects are crushed, torn, dirty, overlapping and partly hidden. A plastic bottle may be flattened into something that looks nothing like a bottle. Labels are missing or covering the wrong thing. The stream never stops and never repeats. Traditional automation - which relies on objects arriving in a known position, one at a time, looking like the thing it was configured for - is helpless here, and for decades the work was done by people standing at the belt.
This turns out to be a good match for machine vision precisely because it is the kind of recognition that does not reduce to rules. A person sorting at a belt cannot explain how they know that particular mangled object is a milk container; they simply recognise it, having seen a great many. Systems trained on very large numbers of images of damaged material work the same way, and they can do it at a rate no person can sustain, for as long as the belt runs. The picking is then done by simple mechanisms - suction cups or air jets - because once the decision is made, the physical part is easy.
The reason this has commercial force rather than being merely clever is that sorting accuracy determines the value of the output. Mixed recyclate is worth very little; a clean, well-separated stream of a single material is worth a great deal more. A facility's economics are therefore extremely sensitive to how well the sorting is done, and to how consistently - human sorters are accurate but variable, particularly late in a shift, and the work is unpleasant, dirty and hard to staff. That combination, a direct link between accuracy and revenue plus a job nobody wants, is the same pattern that drives automation everywhere.
The transferable observation for anyone watching automation generally is about which problems have become tractable. Not long ago, the tasks a machine could do were those that could be specified: place this here, tighten to this torque, reject anything outside these dimensions. Recognition - knowing what a thing is when it does not look like the picture - was firmly human. That boundary has moved, and it has moved in a way that affects far more than recycling. The useful question for any business is which tasks it currently assigns to people purely because they require recognising something, rather than because they require deciding something. Those are different, and only one of them is still safely ours.