Smart Vision-Guided Robotic Depalletising System
Author
Soós, Christoffer Johan
Term
4. semester
Education
Publication year
2026
Submitted on
2026-01-30
Pages
51
Abstract
Automated depalletizing (unloading boxes from a pallet) is difficult because parcels vary widely in size, shape, material, and how they are stacked. Recent AI-based vision has helped robots work in mixed-SKU environments, but robustness, autonomy, and generalization in unstructured conditions are still limited. This thesis designs and implements a robotic depalletizing system that combines camera-based object detection, segmentation, and pose estimation (determining an item's position and orientation). A modular perception pipeline processes RGB-D data (color plus depth) to isolate graspable objects in pallet scenes. The system uses depth-based segmentation with edge-based detection for accurate box identification, and RANSAC-based plane fitting to estimate pose without prior knowledge of parcel dimensions. The system is implemented in a ROS2-based architecture using a UR10 collaborative robot, a VG10 vacuum gripper, and an Intel RealSense D455 RGB-D camera. Experiments across three increasingly complex scenarios achieve 93.3% detection accuracy and an 83.3% overall pick-and-place success rate. These results show strong performance in the tested unstructured settings.
Automatiseret afpalletering (at tage kasser af en palle) er vanskelig, fordi pakker varierer meget i størrelse, form, materiale og i måden, de er stablet. Nye AI-baserede synssystemer har gjort det muligt for robotter at arbejde i miljøer med blandede SKU'er, men der er stadig begrænsninger i robusthed, autonomi og evnen til at generalisere i ustrukturerede situationer. Denne afhandling undersøger design og implementering af et robotsystem til afpalletering, der kombinerer kamerabaseret objektdetektion, segmentering og positursestimering (at bestemme en kasses position og orientering). En modulær perceptionskæde er udviklet til at behandle RGB-D-data (farve- og dybdedata) og finde genstande, der kan gribes, i palle-scener. Systemet bruger dybdebaseret segmentering sammen med kantbaseret detektion til præcis identifikation af kasser, samt RANSAC-baseret plan-tilpasning til at anslå kassernes pose uden på forhånd at kende deres dimensioner. Systemet er implementeret i en ROS2-baseret arkitektur med en UR10 samarbejdende robot, en VG10 vakuumgriber og et Intel RealSense D455 RGB-D-kamera. Eksperimenter i tre scenarier med stigende kompleksitet viser 93,3 % detektionsnøjagtighed og 83,3 % samlet pick-and-place-succesrate. Resultaterne viser god ydeevne i de testede ustrukturerede miljøer.
[This apstract has been rewritten with the help of AI based on the project's original abstract]