Robot Disassembly for Recycling: How It Actually Works

How a Recycling Robot Actually Takes a Product Apart

A recycling robot takes a product apart by following one of two very different playbooks: a fixed sequence built for a single known product, or an adaptive plan built to handle whatever condition the device turns out to be in.

The fixed-sequence approach is already running in production. Apple's disassembly line in Austin, Texas, uses a robot called Daisy that can strip down 23 iPhone models at up to 200 units an hour, roughly 1.2 million phones a year from a single machine. Daisy chills a phone's battery to about -80°C to make it safe to remove, then punches out screws and separates components in around 11 seconds per unit. A companion robot called Dave, added in 2020, opens the Taptic Engine module to recover rare-earth magnets and tungsten, and a shredder-style machine called Taz, added in 2022, separates magnets from audio components without pulverizing them.

That works because every input is the same known product. Daisy's line is engineered around a small, fixed catalog of iPhone models with predictable screw positions, predictable adhesive, and predictable internal layouts. It is fast and reliable precisely because the robot never has to guess what it is looking at.

The adaptive approach solves a different problem: what a robot does when it does not already know the exact condition of the thing in front of it, which describes almost every device that is not a single company's own product line.

The New Approach: Teaching a Robot to Plan Around Damage

A robot handles an unknown or damaged device by comparing it against a CAD model of what it should look like, then updating its plan as it discovers where reality differs from that model.

That is the system a research group at the wbk Institute of Production Science at KIT, in Karlsruhe, Germany, presented at the IEEE International Conference on Robotics and Automation (ICRA) 2026 in Vienna, and which IEEE Spectrum covered in a piece called "How a Smart Recycling Robot Dismantles Old Devices." The lead researcher, Jan Baumgärtner, frames the core problem simply: assembling something new follows a known, fixed sequence, but tearing apart something broken does not, because screws seize, parts go missing, and prior repairs leave a device different from its original design.

The team's paper, "From CAD to POMDP: Probabilistic Planning for Robotic Disassembly of End-of-Life Products," treats disassembly as what researchers call a partially observable decision-making problem. In plain terms, the robot starts with an incomplete picture, a CAD model of the product's design, and has to fill in the missing details, what is actually broken, stuck, or absent, by testing the object as it works.

In practice, the system checks each part's expected degrees of freedom, for example, a screw should turn but not slide sideways, then physically nudges the part to see if it behaves as the model predicts. If a screw turns freely, the robot unscrews it as planned. If a fastener is stuck or corroded, the system switches tools mid-task, using a milling attachment or a grinder to cut through it instead, rather than stalling out on a step a fixed-sequence robot could not adapt around.

The researchers tested the approach on three different products across two robotic systems. Reported results show 30 out of 30 manipulation trials successfully removing the target part, and screw removal succeeded 100 percent of the time when the system added a search step to locate a fastener first, up from 60 percent without it.

Why an Adaptive System Matters for the Circular Economy

An adaptive disassembly robot matters because most electronics recycling has to handle unpredictable, damaged, and mixed products, which a fixed factory line was never built to process. Baumgärtner's stated motivation is to make repair and part-recovery cheaper than manufacturing new components, keeping more devices in a genuine circular economy where they get rebuilt instead of scrapped. His longer-term vision, described in the IEEE Spectrum piece, scales this idea to a factory floor with many coordinated robotic arms working through end-of-life devices at volume, rather than a single research-lab prototype handling one item at a time.

The gap this closes is real and growing. The Global E-waste Monitor 2024, published by the United Nations Institute for Training and Research, found the world generated a record 62 billion kilograms of e-waste in 2022, and just 22.3 percent of that mass was formally collected and recycled in an environmentally sound way. The same report projects e-waste will rise to 82 billion kilograms by 2030, with global recycling rates on track to fall to around 20 percent over that same period if nothing changes.

Dedicated lines like Daisy prove high-volume, high-recovery disassembly works when a company controls its own product design end to end. The harder, larger share of the e-waste stream is everything else: mixed brands, unknown repair histories, and devices that were never engineered to be taken apart by a robot at all. That is the gap adaptive, CAD-driven planning is aimed at closing, though it remains a research result shown at one 2026 conference, not yet a deployed recycling line.

Fixed-Line vs. Adaptive Disassembly, Side by Side

The two approaches trade speed for flexibility in opposite directions, and neither one replaces the other yet.