經由FNN的學習法則來探討機器人動作的相似性

Abstract

學習控制在機器人運動的應用上，雖然學習控制器擁有 generalization 的能力，但是它們卻常被用來輔助傳統控制器。主要的 原因 就是當學習控制器應用在多軸機器人時，由於不同的工作要求產生 多樣的動作，使得學習空間變得很複雜。因此，在這篇論文中，我們首先 針對不同層次的命令做 generalization 能力的探討與比較，找出最適合 做 generalization 的命令。此外，為了降低學習空間的複雜性，我們提 出了動作相似性的分析，使得機器人的動作可根據它們的相似程度予以分 類。在這分析方式中，FNN 先學習各種不同的動作，之後動作間的相似性 可由 linguistic label 的個數以及 membership function 的分佈來判 斷。所以，一組高相似性的動作可用較合理的記憶體大小之學習控制器來 處理，因為這些動作有相似的 FNN 參數，換句話說，也就是有個簡化的 學習空間。 In the application of learning control for robot motion governing, learningcontrollers are usually used as subordinates to conventional controllers, although they are considered to be capable of generalization. One reason is that when a learning controller alone is applied to govern general motions ofmulti- joint robot manipulators, the learning space encountered will be extremely complicated, due to the variations exhibited in motions corresponding to different task requirements. Hence, in this thesis, we first discuss the generalization capability in different levels to find what level command is with the bestgeneralization effect. In addition, in order to reduce the complexity of the learning space for robot learning control, we propose to perform similarity analysis for robot motions by using an FNN learning algorithm, such that robot motions can be classified according to their similarity. In the analysis,the FNN is first used to learn to govern various robot motions, and the similarity between motions is then evaluated according to the number of linguistic labels and the shape of the membership functions of the FNN under successful motion governing. Thus, groups of robot motions with high similarity can be governed by using learning controllers with reasonable sizes, because these motions correspond to similar fuzzy parameters in the FNN, implicating a simplified learning space.