1. Introduction
Municipal recycling facilities increasingly face labour shortages for manual sorting lines, and mixed-material items and contamination make simple colour or weight-based automated sorting insufficient, motivating vision-based approaches capable of discriminating material types under variable lighting and occlusion.
2. Methodology
A YOLOv8-medium detector was fine-tuned on a custom-annotated dataset of 9,600 images covering PET bottles, aluminium cans, cardboard and HDPE containers on a conveyor belt, with detections mapped to world coordinates via a calibrated overhead camera and fed to a six-DOF robotic arm fitted with a vacuum suction gripper for pick-and-place sorting into four bins.
3. Results
The system achieved 94.7 percent mean detection precision at 0.5 IoU across the four material classes, with a successful pick-and-place rate of 89.3 percent, most failures attributable to overlapping items obscuring the gripper approach vector. Sustained throughput averaged 32 items per minute, approaching typical manual sorting-line rates for a single station.
4. Conclusion
Vision-guided robotic sorting using modern single-stage detectors can approach human sorting throughput on well-defined recyclable streams. Future work will address item overlap handling through multi-view camera fusion.
References
[1] Jocher G. et al., YOLOv8, Ultralytics, 2023. [2] Sarc R. et al., Digitalisation and intelligent robotics in value chain of circular economy oriented waste management, Waste Management, 2019.