Instance-aware vision-guided robotic sorting of metal waste using W²G-Net

Abstract

This study aims to develop a robust vision-guided framework for instance-level recognition and grasp-based classification of metal waste in dense, multi-scale and highly reflective industrial environments, addressing the challenges of clutter, mild occlusion and dynamic operation conditions commonly encountered during the post-preprocessing sorting stage of automated recycling systems. A lightweight two-stage What & Where to Grasp framework, termed W²G-Net, is presented. The framework adopts a cascaded pipeline integrating an optimized YOLOv11-Seg for instance-level perception and an improved Generative Residual Convolutional Neural Network (GRCNN) for planar grasp planning. An instance-mask-constrained grasp feasibility modeling method is introduced to enforce geometric and interference constraints, together with a time-compensated dynamic grasp model for conveyor belt scenarios. The approach is evaluated on metal waste segmentation data sets, a newly constructed metal waste 2D grasp data set, and the Cornell grasp data set. Experimental results show that the optimized YOLOv11-Seg achieves a 17.1% inference speed improvement, while the enhanced GRCNN increases inference speed by 59.5%. The proposed two-stage framework operates at 16.2 frames per second (FPS) for single-target and 8.6 FPS for multi-target scenarios, with recognition accuracy and grasp success rates exceeding 90%. Grasp stability is significantly improved in cluttered and dynamic metal-waste environments. This work presents a two-stage grasping framework that combines instance-level segmentation, planar grasp planning and time-compensated dynamic grasp modeling, providing an efficient, scalable and industrially applicable solution for automated metal waste sorting.