Machine learning-driven waste sorting with robotic arms

Abstract

Machine learning-driven waste sorting systems, integrated with robotic arms, can mitigate the limitations of manual sorting. This study develops and deploys a YOLO-based system for automatic identification and sorting of five waste categories: metal, cardboard, glass, paper, and plastic. A key contribution of this study is the integration of the vision-based detection model with robotic arms, implemented in both single-arm and dual-arm experimental setups. To comprehensively evaluate the performance of the purposed system, this study designed a progressive five-layer evaluation framework: from pre-deployment performance of the YOLO model (layer 1) and its adaptability after deployment (layer 2), to the sorting reliability of the single-arm system in both fixed workspace (layer 3) and conveyor workspace (layer 4), and finally, the efficiency and safety validation of the dual-arm system (layer 5). The YOLO model achieved 92.7% precision on the test set before deployment (layer 1). After deployment, precision under static camera views of the robotic arms dropped to 70.9% in the fixed workspace and 65.5% on the conveyor belt (layer 2); single-arm overall success rates were 76.7% (fixed workspace, layer 3) and 63.3% (conveyor workspace, layer 4), and the dual-arm system handled 40% of samples (layer 5), all with 100% picking success for correctly detected objects. The evaluation results show that although object detection performance declines under real-world conditions, especially on dark conveyor belts, robotic grasping remains highly reliable once targets are successfully identified. This confirms that low-cost robotic arms possess stable physical operational capabilities in practica sorting tasks. The study further highlights the limitations of vision-only classification under cost constraints. Given that many small and medium-sized recycling facilities cannot afford industrial-grade equipment, exploring the feasibility and limitations of low-cost solutions holds significant practical relevance. Finally, this study proposes future directions for improvement, including enhancing dataset diversity, integrating multi-sensor, and optimizing dual-arm coordination strategies, with the aim of further improving system efficiency and robustness, thereby providing a reference for promoting economically viable robotic sorting solutions in industrial applications.