Scholarlink

<p>Robotic feeding systems can restore autonomy for individuals with upper limb impairments; however, most existing solutions rely on unimodal perception and rigid control interfaces, which limit adaptability and real-time performance. This study presents a multimodal feeding robot that integrates large-model-based speech understanding, lightweight visual perception, and smooth motion planning. The visual module employs an improved YOLOv8n architecture optimized through a three-stage process of LAMP pruning for compression, fine-tuning for recovery, and BCKD-based knowledge distillation for performance enhancement. Compared with the baseline YOLOv8n, the proposed model achieves improvements of + 1.9% in mAP@0.5 and + 2.0% in mAP@0.5:0.95, while reducing model size from 6.3&#xa0;MB to 2.7&#xa0;MB, computational cost from 8.1 GFLOPs to 3.2 GFLOPs, and parameters from 3.0 million to 1.2 million, resulting in an inference speed increase from 27 to 42 FPS. A 3–5–3 segmented polynomial trajectory ensures smooth insert–lift–transfer motions toward a fixed mouth position, maintaining continuous velocity and acceleration profiles. Across 20 consecutive feeding cycles involving three food types (watermelon, Hami melon, and sausage), the system achieved a 100% success rate, with no failures observed during spearing or delivery. These results demonstrate that the proposed multimodal and lightweight feeding framework enables accurate perception, efficient control, and natural human–robot collaboration, thereby providing a practical pathway toward autonomous assistive mealtime systems.</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

Application of multimodal interaction in robotic feeding assistance for individuals with upper limb disabilities

  • 期刊论文

  • 发表日期：

    2026年2月14日

查看全文
  • Fei Liu,
  • Zhi Li,
  • YuXin Sun,
  • Mingyue Hu

摘要

Robotic feeding systems can restore autonomy for individuals with upper limb impairments; however, most existing solutions rely on unimodal perception and rigid control interfaces, which limit adaptability and real-time performance. This study presents a multimodal feeding robot that integrates large-model-based speech understanding, lightweight visual perception, and smooth motion planning. The visual module employs an improved YOLOv8n architecture optimized through a three-stage process of LAMP pruning for compression, fine-tuning for recovery, and BCKD-based knowledge distillation for performance enhancement. Compared with the baseline YOLOv8n, the proposed model achieves improvements of + 1.9% in mAP@0.5 and + 2.0% in mAP@0.5:0.95, while reducing model size from 6.3 MB to 2.7 MB, computational cost from 8.1 GFLOPs to 3.2 GFLOPs, and parameters from 3.0 million to 1.2 million, resulting in an inference speed increase from 27 to 42 FPS. A 3–5–3 segmented polynomial trajectory ensures smooth insert–lift–transfer motions toward a fixed mouth position, maintaining continuous velocity and acceleration profiles. Across 20 consecutive feeding cycles involving three food types (watermelon, Hami melon, and sausage), the system achieved a 100% success rate, with no failures observed during spearing or delivery. These results demonstrate that the proposed multimodal and lightweight feeding framework enables accurate perception, efficient control, and natural human–robot collaboration, thereby providing a practical pathway toward autonomous assistive mealtime systems.