本研究旨在建構一套整合視覺感知、三維重建、動態手眼校正與避障控制的智慧型機械 手臂系統，以因應智慧製造與人機協作環境中日益提升的彈性與即時性需求。核心技術首先 著重於高效且準確的視覺骨架建立，本研究採用 YOLOv11 Pose 模型對協作型機械手臂進行 即時二維關鍵點偵測，並搭配雙相機立體視覺進行三角測量，重建具結構性的三維骨架。此 三維資訊不僅為後續手眼校正提供基礎，也大幅提升整體系統對手臂動態狀態的掌握能力。 針對實務應用中常見的相機位置偏移問題，進一步導入基於剛體轉換擬合之動態手眼校正技 術，透過將視覺重建骨架與正向運動學推導的理論骨架進行快速對應，持續更新相機與手臂 間的外部參數，避免傳統靜態標定在環境變動下失效的問題，提升系統在動態情境下的穩定 性與精度。在避障控制方面，本系統整合 Intel RealSense D455 深度攝影機，即時擷取周遭點 雲資訊，並透過 DBSCAN 演算法進行雜訊濾除與障礙物建模。於此基礎上，本研究設計兩 種人工勢場避障策略：其一為靜態策略，於任務前利用吸引與斥力場修正目標位置，並結合 逆運動學反覆求解避障路徑，適用於靜態環境；其二為動態策略，於每個控制週期中即時計 算排斥力向量，並經由雅可比矩陣轉換為關節空間修正量，適用於動態或突發障礙物場景。 兩種策略均支援多障礙物處理，增強系統於複雜場域下的靈活性與安全性。;This study aims to develop an intelligent robotic arm system that integrates visual perception, 3D reconstruction, dynamic hand-eye calibration, and obstacle avoidance control, in response to the increasing demand for flexibility and real-time adaptability in smart manufacturing and human-robot collaboration environments. The system first focuses on efficient and accurate skeleton construction by using the YOLOv11 Pose model for real-time 2D keypoint detection of a collaborative robotic arm, combined with dual-camera stereo vision to reconstruct a structured 3D skeleton via triangulation. This 3D data serves as the basis for subsequent hand-eye calibration and significantly enhances the system’s ability to perceive the arm’s dynamic state. To address practical challenges such as camera pose shifts, a dynamic hand-eye calibration method based on rigid-body transformation fitting is introduced. By aligning the vision-based skeleton with the theoretical skeleton derived from forward kinematics, the system continuously updates the external parameters between the camera and the arm, overcoming the limitations of traditional static calibration in changing environments. For obstacle avoidance, an Intel RealSense D455 depth camera is used to capture real-time point cloud data, which is filtered and clustered using the DBSCAN algorithm to build obstacle models. Based on this, two artificial potential field strategies are proposed: a static method that adjusts target positions with attractive and repulsive fields and solves obstacle-free paths using inverse kinematics, and a dynamic method that computes repulsive force vectors in real time and maps them to joint space via the Jacobian matrix. Both strategies support multi-obstacle handling and enhance the system’s flexibility and safety in complex environments.
