ROBNEWS
Industrial robots are projected to quadruple to 16 million by 2030, and the industry has not solved how to take them apart when they break.
There are more than 4 million industrial robots at work in factories today: welders, paint sprayers, pick-and-place arms, the stock population of the modern assembly line. The International Federation of Robotics projects that number will cross 16 million by 2030. Each of those machines will, eventually, fail. The industry has spent decades learning how to build them. It has barely started learning how to take them apart.
Building a robot from new parts follows a script. Each bolt, sensor, and cable has a place, and the steps to put them there are well known. Taking a broken one apart does not. A screw may be seized. A wiring harness may have melted into a housing. A motor may have shifted under years of vibration. The disassembly sequence that worked on the factory floor is useless the moment the machine deviates from its design.
"We can imagine 100 ways that something can go wrong," said Jan Baumgärtner, one of the system's designers, "and the robot doesn't know which one it is."
A team at the Karlsruhe Institute of Technology (KIT) is testing one approach. The group's system starts with a CAD model of the target robot and a probabilistic model that predicts how each part is likely to be damaged. It then attempts disassembly, screw by screw and joint by joint, and checks the result against the prediction at every step. When reality and the model disagree, the system does not stop. It revises. A screw that refuses to come out triggers a switch from unscrewing to milling. A connector that will not release cleanly prompts a different approach to the housing. The loop is predict, act, observe, revise, and the group describes it in a preprint posted to arXiv this month, "From CAD to POMDP: Probabilistic Planning for Robotic Disassembly of End-of-Life Products" (arXiv 2511.23407).
The work is a research artifact from a single design group at KIT's Institute of Production Science (wbk.kit.edu), not a product. The paper is a preprint, not yet peer-reviewed. The question of who pays to put an industrial arm through this kind of careful takedown is, for now, an academic one. A single research group's loop does not solve the industry's end-of-life problem. It shows the loop is buildable.
The IFR's projection is the first hard number to put end-of-life robotics on a fleet-scale footing. By that math, the wave of aging capital equipment is already on the way. The back half of the industrial robot lifecycle is the part the industry has not engineered. The KIT preprint frames the problem in the language a planning researcher would recognize: a partially observable Markov decision process over a CAD-derived part graph.
The trade press has run the work as a clever demo. One robot taking another apart. A stuck screw replaced by a mill. A video that ends before the recycling question begins. The IEEE Spectrum feature that introduced the system walks through the loop in plain language (IEEE Spectrum). A Korean industry pickup at irobotnews.com has covered the same announcement (irobotnews). Both treatments leave the larger question on the table.
The next test is whether anyone outside the lab picks it up. A robot manufacturer's sustainability lead, an industrial recycler, or a circular-economy researcher focused on capital equipment would each have a view on what the second half of an industrial robot's life actually looks like in 2026, and whether a predict-act-observe-revise loop survives contact with a real scrap yard. Until one of those conversations lands, the IFR projection and the KIT preprint point in the same direction without quite meeting: a wave of machines whose afterlife is not yet an engineered process.