Enhancing sorting efficiency in cluttered construction and demolition waste streams via boundary-guided grasp...

Robotic automation is instrumental in the valorization of construction and demolition waste (CDW), facilitating scalable and efficient material recovery in response to rising waste volumes from accelerated urban development. AI-driven computer vision (CV) has advanced perception-focused tasks in CDW processing, such as classification, object detection, and segmentation. However, the subsequent action-focused task of robotic CDW grasp detection remains underexplored. Identifying optimal, collision-free gripper poses for CDW recyclables is particularly challenging in cluttered environments and is further limited by the need for large amounts of grasp-annotated training data. This paper therefore presents a robust CDW grasp detection method that leverages boundary features of CDW objects to guide grasp predictions via attentional feature fusion. Our approach builds upon recent advances in shape-aware CDW instance segmentation and takes advantage of the growing availability of high-quality segmentation data, enabled by automated labelling techniques using large language models (LLMs). To address the shortage of publicly available grasp-annotated CDW data, we also introduce ReCoDeWaste; the first RGB-D CDW instance segmentation and grasp detection dataset tailored for off-site AI-based sorting training and evaluation. Designed to capture the compositional complexity of CDW, ReCoDeWaste contains over 100,000 annotated waste object instances across diverse cluttered scenes. Experimental results demonstrate that our boundary-guided grasp detection model predicts collision-free grasps for detected recyclables in these cluttered streams, outperforming state-of-the-art methods in standard evaluation metrics and achieving up to 94.36% grasp detection accuracy. This research serves to enhance CDW valorization by transitioning efforts to action-focused CV tasks that go beyond recognition and classification.