RAG-BASED ROBOTIC VISION MODEL FOR ADAPTIVE WASTE SORTING AND CLASSIFICATION

Abstract

Effective waste sorting and classification are critical for sustainable waste management; however, current automated systems often lack adaptability and contextual awareness, limiting their accuracy and applicability. We propose a novel robotic vision framework utilizing Retrieval-Augmented Generation (RAG) for adaptive waste classification. This framework integrates real-time visual perception with context-specific knowledge retrieved from external waste management ontology, enhancing object recognition through recyclability, contamination status, and specialized handling insights. Unlike conventional methods dependent solely on pre-trained datasets, our approach dynamically aligns visual data and contextual embeddings to effectively manage complex waste scenarios, including ambiguous or damaged items. Quantitative assessments on a heterogeneous waste dataset revealed the proposed model significantly outperforms traditional methods, achieving an 18% increase in classification accuracy and a 25% improvement in correct recyclability identification. Additionally, our framework markedly reduced hazardous material misclassification, thus enhancing safety in waste management operations. Experimental validations confirm the robustness and flexibility of the RAG-based system, underscoring its practicality for real-world waste sorting applications. By combining intelligent context-awareness with visual adaptability, our robotic vision framework advances sustainable waste classification and automation.