Mitigating Catastrophic Forgetting in Robotic Waste Sorting: A Continual Learning Framework with Dynamic Experience Replay

Abstract

Efficient sorting of critical waste materials (e.g., rare-earth elements, specific plastics, nuclear components) is essential for resource recovery and environmental safety. Deep learning models excel in visual-based sorting but fail to adapt to evolving waste streams without catastrophic forgetting of previously learned knowledge. This paper presents a continual learning framework for robotic waste sorting using classical experience replay. Our approach maintains a balanced memory buffer with reservoir sampling, enabling sequential learning of new material categories while preserving accuracy on previously encountered ones. Experiments on a curated dataset of critical waste materials demonstrate that our method reduces average forgetting compared to fine-tuning baselines and maintains sorting accuracy above 29% on old tasks while integrating new classes with minimal additional data. This work provides a pathway toward adaptive and sustainable waste management systems capable of lifelong operation in dynamic industrial environments.