本論文旨在實作一套基於雙眼視覺回授之桌球機械手臂系統。以Creo 3-D 建模軟體設計5 自由度之手臂以及雙眼攝影機架，並加工實作出來。將左右攝影機影像以桌球顏色二值化再透過輪廓及最小包覆圓法辨識出桌球，最後以立體幾何關係得到球之三維座標進行追蹤。球軌跡點之近似二次曲線用在判斷球是否會打過來並且提供球之初始狀態給飛行模型。飛行模型則被用來預測下一時刻的狀態。預測模型更結合了擴張型卡爾曼濾波器藉由視覺量測修正預測模型以獲得最佳估測狀態。彈跳模型用來預測桌球碰撞到桌子之後在落點反彈起來的初始狀態。依照上述方法，我們可以預測出擊球任務所需要的擊球點。手臂第零軸將使用模糊控制使得手臂能夠左右移動，手臂第一軸使用點到點控制，能夠讓手臂座左右旋轉之動作，剩下三軸則使用函式間補控制，使得手臂能夠上下移動。此系統整合了視覺與控制，能夠將手臂移至預測出的擊球點位置。實驗結果驗證設計之視覺控制系統能有效運作。 This thesis mainly implements a visual feedback control system for table tennis robot arm. The 5-DOF robot arm and binocular camera racks are designed by using Creo 3-D modeling software and then manufactured. The images captured from binocular camera are thresholded according to the ball color then using contour and smallest enclosing circle method to recognize the ball. The 3-D position of the ball is computed by the stereo geometry method. The approximate second-order polynomial of ball trajectory is used to make sure if the ball is coming and it also provides the initial state of the ball for flying model. The flying model is used to predict the state of the ball at the next moment.Extended Kalman Filter is combined to modify the predict result by visual measurement for the best estimation result. The rebound model is used to predict the ball state after rebound at the landing point. By the methods we have mentioned above, we can predict the contact point of the ball for hitting mission. Fuzzy control is implemented in joint 0, which makes the arm move left or right. Profile-position control mode is implemented in joint 1, which makes the arm rotate to hit the ball. Interpolated-position control mode is implemented in left three joints,joint 2 to joint 4, which makes the arm lift up or down. The developed system integrates the vision and control systems together, which can predict the contact point then moves the arm to the right position. Experimental results show that the designed visual feedback control system works effectively.