Scholarlink

<p>To address the challenges of rapid identification and precise localization of threaded hole positions, as well as high-accuracy robot trajectory planning in the automated assembly of automotive wheel hub bolts, this paper presents targeted research. Given the weak texture and high reflectivity characteristics of the wheel hub assembly environment, these factors significantly complicate the tasks of accurate target recognition and pose estimation. Currently, assembly robots typically rely on professionals for teaching through demonstrators or offline programming. This process is not only complex but also limited to simple, specific tasks. As a result, it is insufficient to meet the demands of more intricate product assembly in the era of automation and intelligence. Based on the existing automotive wheel assembly equipment, this paper introduces a 3D vision system and proposes an end-to-end deep learning model for robot trajectory generation in the context of virtual debugging of automobile assembly. First, point cloud data of the assembly environment are captured using the 3D vision system. Then, to prevent the model from overlooking critical information, this paper proposes a novel residual attention mechanism that combines residual connections with a self-attention strategy. This mechanism enhances the expressive power of the network by dynamically weighting features, enabling high-dimensional fusion between local and global representations of the point cloud. Finally, result accuracy is further enhanced through data optimization and the implementation of an effective learning strategy. To validate the reliability of the proposed model, experiments are conducted both in a simulated environment and in a real-world setup. The results demonstrate that the entire assembly process can automatically generate robot trajectories. In the simulated environment, the robot trajectory planning accuracy is improved by an average of 33.27%. In the real assembly scenario, the accuracy is enhanced by an average of 35.99%.</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

AssembleNet: End-to-End Deep Learning Modelling of Robot Trajectories for Virtual Commissioning of Automotive Assembly

  • 期刊论文

  • 发表日期：

    2026年1月5日

查看全文
  • Sheng Wang,
  • Yanzheng Li,
  • Xin Zhang,
  • Yinhua Liu

摘要

To address the challenges of rapid identification and precise localization of threaded hole positions, as well as high-accuracy robot trajectory planning in the automated assembly of automotive wheel hub bolts, this paper presents targeted research. Given the weak texture and high reflectivity characteristics of the wheel hub assembly environment, these factors significantly complicate the tasks of accurate target recognition and pose estimation. Currently, assembly robots typically rely on professionals for teaching through demonstrators or offline programming. This process is not only complex but also limited to simple, specific tasks. As a result, it is insufficient to meet the demands of more intricate product assembly in the era of automation and intelligence. Based on the existing automotive wheel assembly equipment, this paper introduces a 3D vision system and proposes an end-to-end deep learning model for robot trajectory generation in the context of virtual debugging of automobile assembly. First, point cloud data of the assembly environment are captured using the 3D vision system. Then, to prevent the model from overlooking critical information, this paper proposes a novel residual attention mechanism that combines residual connections with a self-attention strategy. This mechanism enhances the expressive power of the network by dynamically weighting features, enabling high-dimensional fusion between local and global representations of the point cloud. Finally, result accuracy is further enhanced through data optimization and the implementation of an effective learning strategy. To validate the reliability of the proposed model, experiments are conducted both in a simulated environment and in a real-world setup. The results demonstrate that the entire assembly process can automatically generate robot trajectories. In the simulated environment, the robot trajectory planning accuracy is improved by an average of 33.27%. In the real assembly scenario, the accuracy is enhanced by an average of 35.99%.