A KIT robot dismantles broken machines to salvage valuable parts

A team at the Karlsruhe Institute of Technology (KIT) showed a robot at the ICRA 2026 conference in Vienna that dismantles broken machines without damaging their valuable parts. The system plans the disassembly order itself, predicts what is broken, and verifies every move. The goal is cheaper repair and part recovery instead of building from scratch.

Key takeaways

The system was presented at IEEE ICRA 2026 in Vienna; the project's designer is Jan Baumgärtner of KIT
The robot unscrews, manipulates, and mills damaged devices while protecting parts the user marks as important
Disassembly planning is framed as a POMDP — a decision process with incomplete information about the product's state
Tested on three products and two robotic systems: shorter and more stable disassembly time than deterministic methods
Over 4 million industrial robots are in use worldwide; the number is projected to top 16 million by 2030

How disassembly under uncertainty works

The starting point is the product's CAD model. From it and from inspection results, the system automatically builds a POMDP model — an extension of the Markov Decision Process that explicitly captures uncertainty. Real, worn-out devices rarely match their documentation: screws seize, parts corrode or go missing.

\dots ▶ Symbol meaning \dots set of possible product states \dots available actions: unscrew, mill, manipulate \dots state-transition probability \dots reward — cost and effect of a move \dots set of inspection observations \dots observation model — what the robot sees in a state \dots discount factor for future rewards

The robot tests a part's degrees of freedomdegrees of freedom: The number of independent directions in which a part can move. to spot the problem — corrosion, a loose screw, deformation. To choose the next move it uses reinforcement learning with stochastic action outcomes, and tracks the product's state with a Bayesian filterBayesian filter: A method that continuously updates a state estimate from new observations and probabilities.. After each step it checks the result and changes its method if reality diverges from the prediction. The user can specify which components matter most and should stay intact.

The robot's decision loop runs through five stages:

CADthe product's starting model

POMDPplanning under uncertainty

DOF testdetect the fault

RL + Bayesian filterchoose a move, update the state

verifycorrect the method after each step

We can imagine 100 ways that something can go wrong.

Jan Baumgärtner, project designer, Karlsruhe Institute of Technology

Probabilistic versus a rigid plan

The authors compared their approach with deterministic methods that assume the product looks exactly like its documentation. Across three products and two robotic systems, the probabilistic model delivered a shorter average disassembly time and lower variance. It also handled deviations a rigid plan does not anticipate — missing or stuck components.

16Mprojected industrial robots worldwide by 2030 (over 4 million today)IEEE Spectrum

Why it matters

Recycling electronics and machines still leans heavily on human labor, because every broken device is different. Framing disassembly as a problem under uncertainty lets a robot act even though it does not know in advance what it will find. That is a step toward profitable part recovery, not just shredding scrap. If repair costs fall below the price of new hardware, the economics of a circular economy change. The growing fleet of well over ten million industrial robots will itself become a large stream of hardware to take apart.

What's next?

The authors want to scale the system to many specialized arms — Baumgärtner describes it as “a giant robot with 100 arms”
The full method is described in an arXiv preprint (From CAD to POMDP), submitted on November 28, 2025
The end goal is repair cheaper than making a new device — a precondition for a real circular economy