Robot learns to dismantle broken machines

Robots have spent decades helping build the products around us. Now researchers are teaching them another skill that could become increasingly important: taking those products apart when something goes wrong.

More than 4.6 million industrial robots are already operating worldwide. Meanwhile, demand for industrial robots continues to grow as manufacturers automate more production. That creates an obvious question. What happens to all those machines and other complex products when parts wear out or fail?

A robotic disassembly system uses probabilistic planning to change tactics when damaged or aging machinery creates unexpected obstacles. (Jan Baumgärtner and research team/KIT)

Researchers at the Karlsruhe Institute of Technology in Germany have developed a robotic disassembly system designed to tackle that problem. Instead of assuming every screw and component will behave perfectly, the system plans for the messy reality of old machines. A screw may be stuck. A component may already be missing. The machine may no longer match the original design. The robot can figure that out as it works and change its plan along the way.

Why broken machines are surprisingly hard for robots

Building something in a factory can be incredibly predictable. A robot knows which part comes next. It knows where the screws belong. Every movement can follow a carefully programmed sequence.

Taking an old machine apart is a very different job. Years of use can leave parts corroded or damaged. Previous repairs can also change how a product fits together. That uncertainty creates a huge problem for traditional automation because one unexpected obstacle can derail the entire disassembly sequence.

Researcher Jan Baumgärtner puts the challenge in practical terms. When assembling something new, the steps are clear. When dismantling something broken, many things can go wrong. That means a robot needs more than instructions. It needs some ability to reconsider what it believes is happening.

How the robotic disassembly system works

The system starts with a CAD model showing how the product should be constructed. From there, the robot examines how individual parts actually behave. It can check whether a component moves the way the model predicts. If the movement looks wrong, the system updates its understanding of the machine. For example, a screw should behave in a very specific way. If the system discovers that a screw moves differently than expected, it can factor that new information into its next decision.

The researchers use a probabilistic planning approach known as a Partially Observable Markov Decision Process, or POMDP. That complicated name describes a fairly relatable idea. The robot knows it does not have perfect information. So, rather than committing to one rigid plan, it assigns probabilities to what might be wrong and keeps updating those assumptions as new information arrives. The research combines that approach with CAD data, inspection information and the capabilities of the robot itself.

Researchers in Germany developed a robotic system that can adapt when screws are stuck, components are missing or machines no longer match their original designs. (Jan Baumgärtner and research team/KIT)

The robot can change tactics when something goes wrong

Here is where this gets interesting. In one physical experiment, the researchers simulated a stuck screw in an electric motor. The robotic system initially tried the expected approach by unscrewing the fasteners. When it discovered that one screw would not cooperate, the robot changed course. Instead of continuing to fight with that screw, the system chose a milling tool to remove material and gain access to the part it wanted. In another test involving an angle grinder, the robot recognized that a screw was already missing and avoided wasting time looking for it.

That adaptability is important because the researchers found that traditional deterministic planning works well when everything behaves exactly as expected. Once uncertainty enters the picture, the probabilistic system can perform better when another disassembly route is available. In the experiments, both approaches performed similarly on new components. As the likelihood of stuck parts increased, the probabilistic planner produced faster disassembly times in scenarios where the robot had another way to reach the target component. The research was presented at the 2026 IEEE International Conference on Robotics and Automation in Vienna.

This robot isn’t tearing apart other robots yet

There is an important distinction here. The researchers are developing technology for robotic disassembly, but the physical demonstrations in this research focused on electric motors and an angle grinder. They did not demonstrate an automated factory where robots dismantle complete industrial robots.

Still, the broader concept could eventually apply to much larger systems. Baumgärtner envisions scaling the technology into facilities with multiple robotic arms equipped with different tools. One machine might handle screws while another tackles parts that require a more aggressive removal method. The long-term vision looks almost like an assembly line running backward.

Could robots eventually make repairs cheaper?

This may be the part you should watch. Baumgärtner says one goal is creating a more circular economy where manufacturers recover useful components from older products instead of discarding the entire device.

The system can even prioritize certain components during disassembly. If a manufacturer says a particular part has significant value, the robot can adjust its strategy to improve the chances of preserving that component. Eventually, the researchers envision an automated process that could extract a bad component, replace it and rebuild the product. Their ultimate economic goal is ambitious: make automated repair inexpensive enough that fixing an electronic device can cost less than producing another one. That remains a goal rather than something the system can deliver commercially today.

The technology could eventually help manufacturers recover valuable components from broken products instead of sending entire machines to the scrap pile. (Jan Baumgärtner and research team/KIT)

What this means to you

You probably will not see one of these robotic repair stations at your neighborhood electronics shop anytime soon. However, this research points toward a different way manufacturers could think about products once they break. Today, many electronics become e-waste because recovering individual components can take too much labor or money. Automation could change some of that math. If robotic systems become good enough at dealing with damaged products, manufacturers could recover more high-value parts.

Refurbishing equipment could also become more economical in some industries. There is another potential benefit. A machine that can intelligently preserve useful components may reduce the amount of perfectly good hardware that gets discarded because one part failed. The big question will be whether manufacturers design future products with automated disassembly in mind. Repair becomes much easier when engineers think about how something will eventually come apart while they are deciding how to build it.

Key takeaways

What catches my attention here is the robot's ability to deal with uncertainty. Factory robots have traditionally thrived in carefully controlled environments where every component arrives in the right place. Broken products refuse to cooperate like that. Teaching machines to recognize when reality no longer matches the blueprint could unlock far more useful applications for robotics. Repair and recycling are especially interesting because the economics often determine whether something gets a second life or lands in the scrap pile. We are still looking at research rather than a repair revolution you can use today. Yet the idea behind it feels important. The smarter robots become at taking products apart, the more realistic it becomes to recover expensive components instead of throwing away an entire machine because one piece failed.