Data-Efficient Deep Learning Methods for Robotic Waste Sorting Systems

The global population increase and the intensifying exploitation of natural resources have highlighted the need for a transition to a more circular economic model. However, current technologies can only partially recover valuable resources from waste due to limitations in efficiency, performance, and economic sustainability. To address these challenges, this thesis focuses on advancing automation in waste sorting plants through the application of cutting-edge artificial intelligence solutions. Specifically, it aims to develop a robotic waste sorting system capable of tackling the key challenges in waste management, with a particular emphasis on data efficiency. A major hurdle in this field is the impracticality of collecting large-scale, real-world datasets, while publicly available data present significant domain shits when compared to industrial settings. Most state-of-the-art deep learning methods rely on extensive training with large datasets, making them less feasible for real-world industrial applications with limited data availability. This thesis investigates various data-efficient techniques to apply deep learning in robotic perception and manipulation, overcoming these constraints. The proposed methods are evaluated using public datasets, self-annotated datasets, and real-world experiments. Additionally, the thesis presents the design and development of a prototype robotic waste sorting system. This prototype simulates the conditions of a waste sorting plant within a laboratory environment, enabling systematic evaluation of the proposed techniques under realistic conditions. The results demonstrate that the developed methods effectively learn from small labeled datasets and can be seamlessly adapted to real-world scenarios. These findings pave the way for practical, data-efficient AI-driven solutions in waste sorting, contributing to the broader goal of sustainable resource recovery and circular economy advancement.