本論文針對小型人形機器人提出一個基於實務型參數最佳化之線上步態訓練系統，主要探討行走步態、雜訊環境下之參數最佳化以及線上步態訓練系統等三大項的設計。在行走步態的設計上，本論文提出一個參數型中樞型態產生器（PCPG），其依據腰部與腳部之末端點的關係來產生人形機器人的行走步態。在雜訊環境下之參數最佳化的設計上，本論文提出兩個最佳化演算法：實務型粒子群最佳化（PPSO）以及實務型全域最佳引導人工蜂群（PGABC）。從實驗結果中顯示，所提出的PPSO與PGABC演算法相較於一般的PSO與GABC演算法確實能夠更有效地選取到一組較佳的參數數值。在線上步態訓練系統的設計上，本論文應用所提出的PPSO以及PGABC演算法來自動選取一組較佳之行走步態參數數值，讓人形機器人可以有效且穩定地走到所指定的目標點。首先設計實現一個步態訓練平台，讓人形機器人能夠依據目前之行走步態的參數值來自動地步行至目標點。當機器人走到目標點時，步態訓練平台會自動地將機器人拉回至起始點，然後繼續執行下一組行走步態參數數值的步態訓練。從實驗結果可知，本論文所提出之PPSO與PGABC演算法可以確實有效地找到一組不錯的行走步態參數數值。 In this dissertation, an online gait training system is proposed for a small-sized humanoid robot based on two practical parameters optimal methods. There are three main designed topics of walking gait, parameters optimization in a noisy environment, and online gait training system. In the walking gait design, a method named Parameter Central Pattern Generator (PCPG) is proposed to generate the walking gait of humanoid robot, and generated by the relationship between waist and feet points. In the parameters optimal design in a noisy environment, two optimal algorithms are proposed: Practical Particle Swarm Optimization (PPSO) and Practical Gbest-guided Artificial Bee Colony (PGABC). Some experiment results are presented to illustrate that the parameter set selected by PPSO and PGABC algorithms is better than that selected by the general PSO and GABC algorithms. In the online gait training system design, the proposed PPSO and PGABC algorithms are applied in the online gait training system to automatically select a better parameter set of walking gait so that the humanoid robot can effectively and stability walk to the assigned goal point. First, a gait training platform is designed and implemented so that the humanoid robot can automatically walk to the assigned goal point according to the current parameters of walking gait. When the robot attains the goal point, the gait training platform will pull it back to the start point. Then the next parameter set will be continuously executed. Finally, some experiment results are presented to illustrate that a better parameter set of walking gait can be efficiently found by the proposed PPSO and PGABC algorithms.
