The last picking station: what a sorting robot changes at the end of the line — and what it does not

The plant in Swisttal-Ollheim is not a sorting hall with one belt but a line that Hündgen built in 2018 “specifically for the clean separation of lightweight packaging waste”. The operator gives the orders of magnitude itself: an annual capacity of 240,000 tonnes, up to 500 tonnes a day, deliveries from around 3 million households. The material runs at 3 metres a second over the conveyors and through screening, air classification, ballistic separation, metal separation, eddy-current separation and near-infrared sorting.

And then the sentence that carries the rest of this text: despite that chain of processes, “still up to 50% sorting residue” is produced.

That is not an admission but a description of what lightweight packaging is: a mixture that machines can separate a long way, but not completely. Everything the units cannot assign cleanly ends up either in the residue or — worse — as a contaminant in a fraction sold as product. The place where the second is prevented is the quality-control stage: the last picking station, where people remove from an almost finished stream whatever does not belong.

In 2022 Hündgen started a project with Bollegraaf Recycling Solutions B.V. of the Netherlands to integrate an AI-guided sorting robot. The reason is one sentence: “To relieve our staff on the belt.” And the place is in the next: Bollegraaf was able to offer a solution “for placing the robot very well in our already existing plant”.

The robot is called RoBB, and the operator describes its task not as sorting but as quality control: “The final phases of the sorting process are decisive for maximising the purity of the waste stream and recovering the desired recyclables — here RoBB can carry out the quality-control tasks with ease, beyond human capabilities.” Bollegraaf names as the concrete objective raising the purity of the PET bottle stream.

The operator's technical details are sparse: a minimum of 50 efficient picks a minute, a patented gripper “that robustly picks up bottles of any shape”, and an AI-guided vision system that identifies and classifies items “based on chemical composition, shape and colour”. Chemical composition here means near-infrared: the robot sees which plastic lies in front of it, not just what it looks like.

What stands out in this description is what is missing. No throughput in tonnes, no purity in percent, no statement on how many people stood at the belt before and after. The operator describes a position and a task. That is less than a brochure promises, and more than most references deliver.

The decision to put the robot at the end follows a calculation that appears on no datasheet and that every plant operator carries in their head.

At the start of a line the stream is mixed. A pick takes one object out of many, and whether it catches the right one decides little, because behind it the screen, the air classifier, the ballistic separator and the near-infrared unit assess the same object again. A missed pick costs one pick. An overlooked object costs nothing, because it travels on.

At the end of the line the stream is almost pure. It has passed through every unit, every tonne has already paid for its processing, and what still lies on the belt is either product or contaminant in the product. A pick here removes a defect from a bale that would otherwise be sold — or rejected. An overlooked object does not travel on; it gets baled.

That is why the question of pick rate is a different one at the last station than at the first. At the front, what counts is how much a machine takes out of the stream. At the back, what counts is how little it overlooks — and how reliably it does so over a shift in which a person no longer has the same attention in the third hour as in the first. The operator says precisely that when it speaks of quality control “beyond human capabilities”: not faster than a person, but more even.

The most remarkable sentence on the operator's page comes before the robot, not after it: “Sorting recyclables by hand on the sorting line is sweaty work and not exactly the most sought-after activity in the recycling process. Nevertheless, this work must continue to be done by human hand.”

A company that has just installed a robot writes, on that robot's page, that hand sorting stays. That is the statement that counts for a procurement. RoBB takes over one station — the quality control of one particular stream, here PET bottles — and the rest of the sorting cabin carries on, with the same people, at the same belts, on the same shift.

For the shift plan that means something concrete. The robot does not relieve the cabin; it relieves one picking station in the cabin. The person who stood there afterwards stands at another — or nobody stands there any more, and the remaining stations are the same as before. Which of the two happens is decided not by the robot but by how many stations the cabin has and which of them handles the stream with the most expensive errors.

And it means something for what to expect of the machine. A plant that still has half its input as residue after ten units does not become a plant with a quarter by adding a gripper at the end. The robot changes the purity of a product, not the yield of the line. Anyone who confuses the two buys the right machine for the wrong metric.

For a person, quality control is the easiest station: the stream is clear, the task is negative — take out what is wrong — and the objects are large. For a robot it is demanding for the same reasons, and the operator's description names the three answers to that.

• The gripper. “Bottles of any shape” means: crushed, with cap, without cap, half full, nested into each other. A gripper that handles only one shape does not overlook things at this station, it grips and misses — and a miss at the end of the line is a contaminant in the bale. The patented gripper is therefore not a footnote but the reason the robot is allowed at this position at all.
• The recognition. Chemical composition, shape and colour. At the last station shape is not enough, because a PET bottle and a PE bottle can look the same. The near-infrared recognition in the robot repeats at the picking station what the plant already did in the unit before — only this time per object, with a pick.
• The installation. The operator stresses that the robot could be placed in the existing plant. A plant from 2018 was not built for a gantry robot over the belt; height, belt width, discharge chutes and maintenance access are given, not plannable. That the integration succeeded in an existing plant is therefore the real reference, more important than the pick rate.

Fitting into the existing plant is also the part that does not stop at acceptance: every delivery changes the stream, and the recognition system has to move with it. Which shares of that ongoing work fall to the operator and which to the supplier is set out for building robots here, and applies no less on a sorting belt.

The Hündgen case is useful because it asks the right question, not because it supplies the answer for a different plant. Four points, in the order in which they should be settled before a quotation.

1. Which stream has the most expensive errors? In Swisttal it was PET. It may be film, paper, beverage cartons. The robot's position follows from that answer, not from the belt layout.
2. What does a rejected bale cost? That is the value of an overlooked contaminant, and hence the value of the consistency the robot delivers. Without that figure the pick rate cannot be translated into money.
3. How many stations does the cabin have, and which one goes? The operator in Swisttal says hand sorting stays. That is probably true everywhere; the only question is which station the robot takes and what happens to the person.
4. Does a robot fit into the existing plant? Height above the belt, access, discharge. The reference proves it worked in a plant from 2018. It proves nothing about one from 2008.

At the last picking station a pilot is comparatively easy to assess, because the purity of the output stream is measured anyway and is the same number before and after the robot. The general traps of a robotics pilot — the manufacturer's engineer standing by, the day shift instead of the night shift — still apply; we have described them separately.

Three caveats.

The source is the operator, and it is brief. Everything quoted here comes from the pages of Hündgen Entsorgungs GmbH & Co. KG and, for the PET objective, from the manufacturer Bollegraaf. There is no independent measurement of purity, availability or staffing effect. The case establishes that a picking robot has been working on quality control in an existing German packaging-waste plant since 2022. It does not establish how well.

The 50 picks are the operator's minimum figure, not a shift measurement. How many of them actually take place on a belt with varying loading and how many of them are correct is not on the page. Both are the real performance figures of a quality-control stage, and both an operator has to measure in their own plant.

It is a lightweight-packaging plant. The stream, the object sizes and the error costs of a packaging line differ from those of commercial-waste or construction-waste processing. The logic of the last position applies everywhere; the picks needed there do not. Anyone taking something away from Swisttal takes the question, not the machine.