A grasp point generation algorithm for waste handling based on a generative reasoning network

RESEARCH ARTICLE A grasp point generation algorithm for waste handling based on a generative reasoning network Xiao Xiao1, Deyu Liu1, Hai Qin 2,3 * 1 School of Electrical Engineering, Hunan Industry Polytechnic, Changsha, Hunan Province, China, 2 School of Electronic Information, Hunan First Normal University, Changsha, Hunan Province, China, 3 Key Laboratory of Hunan Province for 3D Scene Visualization and Intelligence Education, Changsha, Hunan Province, China * Abstract OPEN ACCESS Citation: Xiao X, Liu D, Qin H (2026) A grasp point generation algorithm for waste handling based on a generative reasoning network. PLoS One 21(5): e0349864. https://doi.org/10.1371/ journal.pone.0349864 Editor: Marco Antonio MorenoArmendariz, Instituto Politecnico Nacional, MEXICO Received: November 15, 2025 Accepted: May 6, 2026 Published: May 29, 2026 Copyright: © 2026 Xiao et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data availability statement: All relevant data are within the paper. Funding: This work was supported in part by the Project of the Hunan Provincial Natural Science Foundation (2026JJ90097), the Scientific Research Fund of Hunan Provincial Education Department (25A0672), and the In the process of urban kitchen waste sorting, robots often encounter issues such as slipping and empty grabs when attempting to grasp dirty waste objects like plastic bottles and glass bottles. This paper proposes a garbage grasping framework based on the Channel Exchange Generative Residual Inference Network (CE-GR-NET), which synthesizes optimal grasping trajectories through the fusion of hierarchical visual features. The object detection network identifies and locates recyclable bottles among solid waste, while CE-GR-NET uses RGB and depth images to generate grasping points for plastic recyclable bottles. Experimental results show that, on the Cornell Grasping Dataset, the proposed method achieves good inference performance and fast inference speed in single-object solid waste scenarios, ultimately generating grasping boxes for plastic recyclable bottles in RGB images. On the self-constructed multi-source urban kitchen waste images, the proposed method generates grasping boxes for targets and regresses the corresponding object categories simultaneously, achieving an image-based model accuracy of 96.03%, an objectbased model accuracy of 94.40%, and a grasping object classification accuracy of 97.87%. Introduction Waste management is a global challenge, with hundreds of millions of tons of waste generated daily worldwide, requiring significant resources and costs for proper handling. According to current data and projections, from 2020 to 2050, municipal solid waste is expected to emit approximately 32–35 billion tons of carbon dioxide equivalent greenhouse gases [1,2]. Due to rapid urban population growth and the limited capacity of urban environments, municipal solid waste (MSW) often cannot be properly managed, leading to increasingly serious environmental problems. This poses a PLOS One | https://doi.org/10.1371/journal.pone.0349864 May 29, 2026 1 / 18 Research and Practice of the “Dual Education” Talent Cultivation Model Based on Vocational Skill Level Certificates (YB2020010102). Competing interests: The authors have declared that no competing interests exist. significant obstacle to sustainable development. Among urban MSW, there are recyclable components such as plastic bottles, aluminum cans, and Tetra Paks. Waste recycling plays a vital role in reducing emissions, creating value, lowering operational costs, and generating both social and economic benefits. Currently, manual waste sorting faces challenges such as high labor intensity and low efficiency. In addition, traditional image recognition algorithms often rely on limited and homogeneous datasets, which are expensive to construct and lack generalizability. Consequently, the stochastic nature of waste orientation necessitates a more robust framework for pose estimation and grasping accuracy [3]. In recent years, with the rapid development of deep learning techniques, semisupervised learning has emerged as an effective approach to address challenges such as the high cost of data annotation and the scarcity of labeled data. Among various semi-supervised learning methods, generative models [4] play a significant role by combining generative and discriminative paradigms. These models leverage both labeled and unlabeled data to enhance model performance while generating realistic data samples.In practical applications such as waste classification, grasp point generation is a critical task that can significantly improve the efficiency of robotic waste sorting. To address this problem, Kumra et al [5]. proposed GR-ConvNet (Generative Residual Convolutional Neural Network), which integrates image analysis and robotic control. Through deep learning, GR-ConvNet is capable of generating accurate grasp rectangles from images, providing essential support for robotic systems to grasp waste objects in complex environments. This semi-supervised generative model learns from unlabeled waste images and generates precise grasp frames. Urban kitchen waste, in addition to biodegradable biomass, often contains a large proportion of other household waste such as plastics, glass, fibers, and metals. Among these, “bottle-like” waste—including plastic bottles, Tetra Paks, aluminum cans, and glass bottles—tends to have relatively intact shapes and visually similar appearances. These items not only require high detection precision but also possess high recycling value. This paper focuses on bottle-like waste as the primary research object and explores grasp rectangle generation using visual methods. Currently, most research on grasp inference does not involve the classification or identification of the target objects. In this work, we propose CE-GR-NET, a grasp rectangle generation algorithm based on a semi-supervised generative reasoning network. Built upon GR-ConvNet, the proposed model integrates a channelexchange module and combines object detection and grasp inference networks. This enables the system to perform targeted grasp inference on recyclable objects such as plastic bottles with higher purposefulness and improved sorting efficiency. Related work Recent research has made considerable progress in the recognition and classification of waste images [6]. Tachwali et al. proposed a classification method based on thermal imaging, utilizing Support Vector Machines (SVM) to detect three different types of recyclable items from thermal images. In addition, a decision tree was employed to classify plastic bottles based on chemical composition and color, PLOS One | https://doi.org/10.1371/journal.pone.0349864 May 29, 2026 2 / 18 achieving an ac (...truncated)