RTDRNet-lite: A lightweight real-time detection framework for robotic waste sorting.

Abstract

In the age of global urbanization, waste recycling remains a critical challenge, impacting the environment and societies from small communities to entire nations. This research aims to address these gaps by proposing a comprehensive and fully automated waste management framework that integrates advanced AI-based detection with robotic hardware to enable intelligent, real-time waste sorting. The fundamental framework of this work is the RTDRNet-lite model, a modified lightweight version of the high-performing object detection variant RT-DETR, which achieved an impressive mAP@50 of 97%. Developed with real-time applicability in mind, the model uses lightweight C2F modules within its head architecture, reducing the computational complexity without any dramatic change in accuracy. A unique approach to training the model was employed, leveraging both real-world waste image data and highly detailed synthetic images generated using the Stable Diffusion model, the Realistic Vision v5.1. This hybrid approach enriches visual diversity and improves the model's generalizability, especially in handling complex object boundaries. The model is trained on four high-frequency waste categories, paper, plastic, glass, and metal, using over 12,929 annotated instances. Additional qualitative evaluations, including IoU-based visual analysis, external validation, and heatmap visualization, confirm the model's robustness, spatial accuracy, and resilience in complex scenes. To demonstrate real-world applicability, a custom 4-degree-of-freedom (DoF) robotic arm was developed and integrated with the model, successfully validating its performance in live sorting tasks. The results confirm both the numerical performance and the practical deployment potential of the proposed system for large industrial-scale waste management facilities and environments.