本論文主要研究內容為利用模糊邏輯結合類神經網路而成的適應性類神經模糊推論系統ANFIS（Adaptive Neuro–Fuzzy Inference System）取代傳統上，使用預設或單純使用模糊邏輯產生ZMP值的方法。其方法是利用理想資料對於ANFIS系統進行學習訓練，建立一個輸入-輸出映射關係，便可利用二足機器人運動時，各部位的幾何位置求解出穩定步行的ZMP值，進ㄧ步利用軀幹補償來達到穩定步行的目的。在模擬方面，建立一個虛擬八連桿二足機器人配合MSC.VISUAL NASTRAN 4D 與 MATLAB SIMULINK、FUZZY Toolbox 進行設計與即時運算模擬，利用直線步行、行進間變換速度和斜坡加以測試，以驗證本文所提之方法可以提供機器人穩定步行並提供不同於以往對於ZMP值的設置方法。 The zero-moment-point (ZMP) trajectory in the biped robot’s foot is a significant criterion for the stability during walking. Many related methods have been proposed for ZMP trajectory. This thesis proposes an Adaptive-Network-based Fuzzy Inference System (ANFIS) ZMP trajectory generator that would set ZMP trajectory more easily and efficiently. By using a hybrid learning procedure and human knowledge input-output data pairs, the ANFIS system can construct a set of fuzzy if-then rules with appropriate membership function to generate the input-output mapping relationship. The hybrid learning procedure combines the gradient method and the least squares estimate. We can use the robot’s hip and leg trajectory as input data. ANFIS generates stable ZMP values as output data. Base on the stable ZMP values, the trunk was used to provide compensation for robot stable walking. The proposed method, under the simulation conditions, can let the biped robot walk stably. The ZMP generated in this method different with the previous method in the literature. In simulation, an eight links biped robot was applied. A real-time simulation in walking a straight line, changing velocity and on slop surface was conducted by using several software such as MSC.VISUAL NASTRAIN, MATLAB ANFIS Toolbox, and MATLAB SIMULINK etc.