Scholarlink

<p>Robotic bin-picking research has long focused on computer vision for object recognition and grasp planning. However, much of the existing work is still based on the oversimplified experimental settings with minimal clutter and occlusion, reducing robustness in real industrial environments. Adoption is further constrained by the difficulty in autonomously handling textureless or highly reflective components, as well as the scarcity of large, high-quality datasets required for deep-learning-based detection and localization. This study presents a fully automated pipeline for bin-picking of textureless parts in highly cluttered scenes, enabled by a deep-learning-based instance segmentation model trained entirely on synthetic images. We introduce a novel synthetic data generation algorithm tailored for instance segmentation, followed by a grasp planning framework based on a 3D reconstruction of the bin environment obtained from an RGB-D camera. The integrated bin-picking system was evaluated using a standard industrial robot manipulator. Our method achieved a 90% grasp success rate in highly occluded environment, demonstrating the practical viability of synthetic data for real-world bin-picking. Experimental validation confirms the system’s effectiveness in accurate object recognition and reliable grasping, offering improved efficiency and adaptability for industrial automation.</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

Occlusion-resilient pose estimation of textureless components in cluttered environment and its implementation in robotic bin-picking

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Nitin Desai,
  • Debashis Sen,
  • Sankha Deb

摘要

Robotic bin-picking research has long focused on computer vision for object recognition and grasp planning. However, much of the existing work is still based on the oversimplified experimental settings with minimal clutter and occlusion, reducing robustness in real industrial environments. Adoption is further constrained by the difficulty in autonomously handling textureless or highly reflective components, as well as the scarcity of large, high-quality datasets required for deep-learning-based detection and localization. This study presents a fully automated pipeline for bin-picking of textureless parts in highly cluttered scenes, enabled by a deep-learning-based instance segmentation model trained entirely on synthetic images. We introduce a novel synthetic data generation algorithm tailored for instance segmentation, followed by a grasp planning framework based on a 3D reconstruction of the bin environment obtained from an RGB-D camera. The integrated bin-picking system was evaluated using a standard industrial robot manipulator. Our method achieved a 90% grasp success rate in highly occluded environment, demonstrating the practical viability of synthetic data for real-world bin-picking. Experimental validation confirms the system’s effectiveness in accurate object recognition and reliable grasping, offering improved efficiency and adaptability for industrial automation.