Scholarlink

<p>To address the challenges of grasping cluttered fasteners and the high cost of manual annotation, this paper proposes a deep learning-based cascaded grasping model and a synthetic data generation pipeline for fasteners. The cascaded grasping model first detects graspable fasteners using an improved YOLO v8n, followed by an estimation of the optimal grasping pose using an enhanced Generative Residual Convolutional Neural Network (GRCNN). The synthetic data generation pipeline leverages physics-based simulation with domain randomization to create synthetic cluttered fastener datasets, thereby augmenting the real dataset and effectively reducing the reliance on manual annotation while mitigating the generation of non-independent and identically distributed (non-IID) data. Experimental results demonstrate that the generated synthetic data significantly enhances model performance, and the proposed improvements effectively boost detection accuracy and robustness. In real-world grasping tasks, the proposed cascaded grasping model achieves an average grasping success rate of 92% across different fastener types and varying levels of clutter.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Robot arm grasping for cluttered fasteners based on deep learning with synthetic data augmentation

  • 期刊论文

  • 发表日期：

    2025年10月7日

查看全文
  • Yiliang Huang,
  • Qiong Yan,
  • Haijun Zhang,
  • Junjie Wang,
  • Quanwei Liu,
  • Qingyu Li

摘要

To address the challenges of grasping cluttered fasteners and the high cost of manual annotation, this paper proposes a deep learning-based cascaded grasping model and a synthetic data generation pipeline for fasteners. The cascaded grasping model first detects graspable fasteners using an improved YOLO v8n, followed by an estimation of the optimal grasping pose using an enhanced Generative Residual Convolutional Neural Network (GRCNN). The synthetic data generation pipeline leverages physics-based simulation with domain randomization to create synthetic cluttered fastener datasets, thereby augmenting the real dataset and effectively reducing the reliance on manual annotation while mitigating the generation of non-independent and identically distributed (non-IID) data. Experimental results demonstrate that the generated synthetic data significantly enhances model performance, and the proposed improvements effectively boost detection accuracy and robustness. In real-world grasping tasks, the proposed cascaded grasping model achieves an average grasping success rate of 92% across different fastener types and varying levels of clutter.