Vision-Guided Robotic Sorting System for UV-Based Magnesite Mineral Classification
Author
Avgeropoulos, Konstantinos Panagiotis
Term
4. semester
Education
Publication year
2026
Submitted on
2026-06-05
Abstract
This thesis explores a laboratory proof of concept where a robot uses UV-C light to sort small magnesite stones from mixed mineral waste streams. The basic idea is that magnesite glows differently than other minerals under UV-C illumination, and that a robot can use this effect to sort the stones automatically. The proposed system combines several steps: UV-C fluorescence imaging, a computer vision method (YOLOv2-based instance segmentation) to detect individual stones, color and brightness analysis in the CIELAB color space, multi-object tracking over time, prediction of where and when a stone can be intercepted, and a robot arm that pushes stones instead of gripping them. The end-effector (the part of the robot that touches the stone) is made from a flexible plastic material (TPU) that can deform on impact. The system was evaluated in two separate experiments: one for perception (the “seeing” part) and one for actuation (the movement and pushing). Under UV illumination, the perception pipeline was tested on 200 labelled rocks and achieved 76.00% accuracy and a 63.64% F1-score, showing that UV fluorescence can support distinguishing magnesite from other minerals. Under full illumination, the tracking, timing, and pushing pipeline was tested: it achieved 167 successful pushes out of 200 attempts, corresponding to an 83.5% success rate. This demonstrates that timed, non-grasping interception of small stones is mechanically feasible in this laboratory setup. However, the UV-based classification and robotic pushing were not tested together as one fully integrated UV sorting system. The thesis should therefore be viewed as a partial validation of the necessary sensing and actuation principles, rather than a complete demonstration of a working industrial sorter. Future work should focus on more stable UV illumination, improved imaging hardware, more powerful computation, and fully integrated end-to-end testing where the entire system runs in one continuous process.
Dette speciale undersøger i et laboratorieforsøg, om det er muligt at lade en robot sortere små magnesit-sten fra blandede mineralaffaldsstrømme ved hjælp af UV-C-lys. Ideen er, at magnesit lyser anderledes end andre mineraler, når det belyses med UV-C, og at en robot kan udnytte denne forskel til automatisk sortering. Systemet kombinerer flere trin: UV-C-fluorescensbilleder, et billedanalyseværktøj (YOLOv2-baseret segmentering) til at finde de enkelte sten, farve- og lysstyrkeanalyse i CIELAB-farverummet, sporing af flere objekter over tid, beregning af hvor og hvornår stenen vil være til at ramme den, samt en robotarm, der skubber til stenene i stedet for at gribe dem. Endeeffekten (den del af robotten, der rører ved stenen) er lavet af et blødt plastmateriale (TPU), som kan give efter ved kontakt. Systemet blev testet i to adskilte forsøg: ét for ”synet” (perception) og ét for robotbevægelsen (aktuation). Under UV-belysning blev syns-delen testet på 200 mærkede sten og opnåede 76,00 % nøjagtighed og en F1-score på 63,64 %, hvilket viser, at UV-fluorescens kan bruges til at skelne magnesit fra andre mineraler. Under fuld belysning blev sporing, timing og skub af sten testet: her lykkedes 167 ud af 200 skub, svarende til en succesrate på 83,5 %. Det viser, at det er mekanisk muligt for robotten at ramme og skubbe små sten korrekt i denne laboratorieopsætning. De to dele – UV-baseret klassifikation og robotisk skub – blev dog ikke testet som ét samlet, sammenhængende UV-sorteringssystem. Specialet skal derfor ses som en delvis dokumentation af, at de nødvendige sanse- og bevægelsesprincipper kan fungere, men ikke som et fuldt bevist, praktisk sorteringsanlæg. Fremtidigt arbejde bør fokusere på mere stabil UV-belysning, bedre kameraer, stærkere computerkraft og samlede end-to-end tests, hvor hele systemet kører i én samlet proces.
[This abstract has been rewritten with the help of AI based on the project's original abstract]
Keywords
robotics ; computer-vision ; minerals ; magnesite ; industrial ; conveyor belt ; yolo ; otsu ; cielab