Scholarlink

<p>To ensure collision safety between humans and robots in collaborative robot environments, it is essential to predict collision forces and pressures in advance based on various end-effector geometries. Since existing geometric formulas are limited to simple shapes, this study proposes a new dual-branch-point transformer framework capable of predicting collision outcomes for objects with arbitrary shapes. The proposed network directly processes 3D point clouds to extract hierarchical geometric features and combines these with dynamic collision parameters (velocity, mass, and angle) to estimate maximum collision force and pressure. Experimental results demonstrate that the proposed model achieves an inference time of 50 ms, enabling real-time safety assessment. Quantitatively, it recorded a mean percentage error (MPE) of 16.8 % for impact force and 14.2 % for pressure, representing a performance improvement of approximately 65 % compared to existing 3D CNN-based methods. Furthermore, the model’s high stability and generalization performance across various collision scenarios were confirmed through the RMSE metrics (8.42 N and 1.22 MPa) and regression line slopes of 0.91 and 0.96. This study provides a practical and robust tool for pre-verifying compliance with ISO/TS 15066 safety regulations in dynamic industrial environments.</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

Method for improving collision force/pressure prediction performance between human and robot colliders using an arbitrary shape point cloud-based transformer

  • 期刊论文

  • 发表日期：

    2026年8月5日

查看全文
  • Younghoon Cha,
  • Heonseop Shin,
  • Sungsoo Rhim

摘要

To ensure collision safety between humans and robots in collaborative robot environments, it is essential to predict collision forces and pressures in advance based on various end-effector geometries. Since existing geometric formulas are limited to simple shapes, this study proposes a new dual-branch-point transformer framework capable of predicting collision outcomes for objects with arbitrary shapes. The proposed network directly processes 3D point clouds to extract hierarchical geometric features and combines these with dynamic collision parameters (velocity, mass, and angle) to estimate maximum collision force and pressure. Experimental results demonstrate that the proposed model achieves an inference time of 50 ms, enabling real-time safety assessment. Quantitatively, it recorded a mean percentage error (MPE) of 16.8 % for impact force and 14.2 % for pressure, representing a performance improvement of approximately 65 % compared to existing 3D CNN-based methods. Furthermore, the model’s high stability and generalization performance across various collision scenarios were confirmed through the RMSE metrics (8.42 N and 1.22 MPa) and regression line slopes of 0.91 and 0.96. This study provides a practical and robust tool for pre-verifying compliance with ISO/TS 15066 safety regulations in dynamic industrial environments.