中国科学院自动化研究所机构知识库

随着机器人与人工智能技术的不断发展，机器人在众多工业领域得到了广泛应用，用以替代人工作业，降低了工人的劳动强度和作业风险，提高了作业质量。但在一些复杂、危险的作业环境中，尤其是在空间狭窄、作业受限的环境中，作业人员和传统的工业机械臂很难甚至无法进入作业空间内，严重制约了特种场景下作业效率和作业质量的提升。本文针对复杂狭小环境下的涂胶作业问题，就超冗余度机械臂涂胶作业系统的设计、运动规划、末端定位和视觉检测方法展开研究，主要内容如下：

第一，介绍了本文的研究背景及意义，综述了超冗余度机械臂的结构设计、运动规划、末端定位与目标检测方法的研究现状，给出了本文的主要内容和结构安排。

第二，面向复杂狭小空间内涂胶作业的任务需求，研制了一款绳索驱动的蛇形机械臂涂胶作业系统，具备超冗余自由度的结构特点，并通过机械臂末端安装的涂胶作业工具和视觉系统，使其可以灵活避开涂胶场景内的障碍物，到达涂胶作业位置开展涂胶作业，并实时检测涂胶作业质量。

第三，针对超冗余度蛇形机械臂的运动规划问题，提出了一种多约束条件下的两阶段运动规划方法，第一阶段根据环境约束和作业姿态约束规划路径曲线，第二阶段根据机械臂形体约束规划机械臂姿态。针对机械臂在高次路径曲线方程下的姿态求解问题，提出了一种基于曲线拟合的数值迭代方法。在单路径的穿越运动任务中，根据蜷缩式推送机械臂的结构特点进行了算法加速，同时探讨了使用神经网络进行算法加速的可能性。算法应用到了不同推送形式的超冗余度蛇形机械臂，使其能够实时跟随目标路径运动，并在运动过程中保持平滑的形体姿态。

第四，针对涂胶作业的作业点识别和定位问题，构建了基于结构光相机的视觉系统，设计了一种基于几何筛选和区域权重聚类的胶缝特征点提取算法，能够稳定筛选并提取胶缝特征点。经过多场景实验测试，初步验证了算法的有效性。通过结构光视觉系统的标定，可以将二维胶缝特征点图像坐标转化为世界坐标系下的三维坐标。经样机测试，定位精度能达到0.1mm。此外，针对涂胶作业后的胶缝质量检测问题，通过样本采集和数据增广的方式构建了胶缝缺陷数据集。以深度学习目标检测算法Yolov7为基本框架，通过聚类得到了合适的初始检测锚框，针对样本多样性不足和数据长尾分布问题对样本采取了增广和降采样措施，同时使用 focal loss替换了交叉熵损失。最终提高了网络模型在新场景下的胶缝质量检测能力并成功应用于该场景。

最后，对本文的工作进行了总结，并指出了可以进一步开展的工作。

With the continuous development of robots and artificial intelligence technology, robots have been widely used in many industrial fields to replace manual work, thereby reducing the labor intensity and operational risks for workers while enhancing operational quality. However, in some complex and dangerous operating environments, especially in environments with narrow spaces and restricted operations, it is difficult or even impossible for operators and traditional industrial robot arms to enter the operating space, which severely restricts the enhancement of operational efficiency and quality in specialized scenarios. Aiming at the problem of gluing operations in complex and narrow environments, this paper conducts research on the motion planning, end positioning and visual inspection methods of a 
hyper-redundant snake-shaped manipulator. The main contents are as follows:

Firstly, the research background and significance of this paper are introduced. The current research status of structural design, motion planning, end positioning and target detection methods of hyper-redundant manipulators are reviewed. And the main content and structural arrangement of this paper are given.

Secondly, in order to meet the task requirements of gluing operations in complex and small spaces, a cable-driven snake-shaped manipulator gluing operation system was developed. It possesses the structural characteristics of hyper-redundant degrees of freedom and is equipped with an adhesive application tool and a visual system mounted at the end of the robotic arm. This enables it to flexibly navigate around obstacles within the adhesive application scene, reach the adhesive application position, conduct adhesive application, and perform real-time detection of adhesive application quality.

Thirdly, addressing the motion planning problem of hyper-redundant serpent-shaped robotic arms, a two-stage motion planning method under multiple constraints is proposed. In the first stage, a path curve is planned based on environmental constraints and operational posture constraints; in the second stage, the arm posture is planned based on the constraints of the robotic arm's form. To address the posture solving problem of the robotic arm under high-degree path curve equation, a numerical iterative method based on curve fitting is proposed. In single-path traversal motion tasks, algorithm acceleration is performed based on the coiling-push structure of the robotic arm, while also exploring the potential of utilizing neural networks for algorithm acceleration. The algorithm is applied to hyper-redundant serpent-like robotic arms with different pushing forms, enabling them to follow target paths in real-time while maintaining smooth body posture during motion.

Fourthly, addressing the issue of adhesive application point recognition and localization, a vision system based on structured light cameras was constructed. A adhesive seam feature point extraction algorithm, based on geometric filtering and region-weighted clustering, was designed. This algorithm is capable of stable selection and extraction of adhesive seam feature points. The effectiveness of the algorithm was preliminarily verified through multiple-scene experimental tests. Through calibration of the structured light vision system, two-dimensional adhesive seam feature point image coordinates can be transformed into three-dimensional coordinates in the world coordinate system. Prototype testing demonstrated a positioning accuracy of up to 0.1mm. Furthermore, to address the issue of glue seam quality inspection after gluing operations, a glue seam defect data set was constructed through sample collection and data augmentation. Utilizing the Yolov7 deep learning object detection algorithm as the fundamental framework, appropriate initial detection anchor boxes were obtained through clustering. Measures such as augmentation and downsampling were employed to address the issues of insufficient sample diversity and long-tailed data distribution. Additionally, focal loss was utilized to replace cross-entropy loss. These efforts ultimately enhanced the network model's ability to detect adhesive seam quality in new scenarios, leading to successful application.

Finally, the conclusions are given and future work is listed.