Scholarlink

<p>Accurate part segmentation in real assembly camera views is important for safe robotic manipulation in confined manufacturing workspaces, where perception errors can affect grasp selection, collision avoidance, and motion planning in environments shared with fixtures and human operators. However, segmentation often degrades under occlusion, clutter, and difficult lighting, where thin boundaries and partially visible parts are common. In many practical assembly, packaging, and intralogistics settings, additional 3D sensing or proprietary CAD assets are not always available, motivating vision-only methods that leverage accessible structural priors. We present Manual-Guided Assembly-Part Segmentation (MAPS), which investigates whether step-aligned isometric assembly manuals can improve camera-view part segmentation without requiring 3D assets as model inputs or at deployment time. Built on YOLO11s-seg, MAPS uses a dual-input teacher-student design in which a frozen manual-trained teacher provides diagram-derived guidance at inference, while a manual-guided auxiliary-head prior filters detector instance masks to form a refined union mask. We evaluate MAPS on the IKEA-Manuals-at-Work dataset through quantitative comparison, qualitative overlays, manual-prior reliability analysis, ablations, unseen-category generalization, and shop-floor difficulty stratification. MAPS improves union-mask accuracy over a camera-only baseline, increasing Dice by 10.4% and IoU by 16.7%, while improving BoundaryF from 0.3043 to 0.5815. Gains persist across difficulty conditions while maintaining practical task-level inference latency. Reliability analysis shows that performance depends on structural compatibility between the selected manual prior and the visible part configuration. Overall, isometric manuals provide a low-cost structural cue for robust visible assembly-part segmentation, motivating future work on reliability-aware manual selection and downstream robot manipulation.</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

Design and evaluation of an assembly part segmentation framework using 2D structural guidance for robotic manipulation in confined workspaces

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月23日

查看全文
  • Chia Ying Kuo,
  • Sheng Jen Hsieh

摘要

Accurate part segmentation in real assembly camera views is important for safe robotic manipulation in confined manufacturing workspaces, where perception errors can affect grasp selection, collision avoidance, and motion planning in environments shared with fixtures and human operators. However, segmentation often degrades under occlusion, clutter, and difficult lighting, where thin boundaries and partially visible parts are common. In many practical assembly, packaging, and intralogistics settings, additional 3D sensing or proprietary CAD assets are not always available, motivating vision-only methods that leverage accessible structural priors. We present Manual-Guided Assembly-Part Segmentation (MAPS), which investigates whether step-aligned isometric assembly manuals can improve camera-view part segmentation without requiring 3D assets as model inputs or at deployment time. Built on YOLO11s-seg, MAPS uses a dual-input teacher-student design in which a frozen manual-trained teacher provides diagram-derived guidance at inference, while a manual-guided auxiliary-head prior filters detector instance masks to form a refined union mask. We evaluate MAPS on the IKEA-Manuals-at-Work dataset through quantitative comparison, qualitative overlays, manual-prior reliability analysis, ablations, unseen-category generalization, and shop-floor difficulty stratification. MAPS improves union-mask accuracy over a camera-only baseline, increasing Dice by 10.4% and IoU by 16.7%, while improving BoundaryF from 0.3043 to 0.5815. Gains persist across difficulty conditions while maintaining practical task-level inference latency. Reliability analysis shows that performance depends on structural compatibility between the selected manual prior and the visible part configuration. Overall, isometric manuals provide a low-cost structural cue for robust visible assembly-part segmentation, motivating future work on reliability-aware manual selection and downstream robot manipulation.