類神經網路動態估測器與機械臂適應性力/位置控制器設計與實驗

Abstract

本論文首先針對順應性運動控制及其不確定性的問題提出一個卡式標 下基於模型的混合式控制器設計。對於模型的混合控制器雖然可以解決傳 統控制器高度非線性,力矩耦合,穩定度無法證明的限制。但是,控制器本 身需要準確的受控體的數學模型;然而,在一般的情形下,我們只能估測系 統的動態方程式,而無法得到準確的模型。這部份我們利用加強式學習的 設計來加以解決。使得基於模型的混合式控制器能夠達到我們預期的表現 。文中並以一個二軸機器臂為例,在事先不知道其動態模型的情況下,經過 重複的學習而逐漸找到最佳的表現,達成同時控制力量與位置的目的。我 們以電腦模擬及實驗來驗證所發展的力/位置控制器設計確能達成預期的 目標。 This thesis addresses the problem of complaint motion control of robotmanipulators. We propose a new adaptive control scheme to deal with the uncertainties of complaint motion. An improved resolved accerlation Cartesian hybrid controller based on the dynamic model of robot manipulators is proposed. The restrictions in conventional design such as nonlinearity, torquecoupling, and the proof of stability can be solved using this model-based hybrid control approach. The controller design requires the exact mathematicalmodel. However, in general case, one can only estimate the system equations, but can not obtain the exact dynamic model. Therefore, we further propose a neural network dynamic estimator to overcome the problem of inexact dynamicmodel. This design allows model-based hybrid controller achieve the predictedperformance. To demonstrate the performance of this control design, we apply reinforcement learning control on a two-link robot arm. Simulation results show that the optimal performance for force and positioncan be obtained simultaneously. Practical experiments on a self- constructed two-link direct drive robot arm is presented to show the possibility of applying this control scheme in real world tasks.