袋鼠機器人之動態分析和狀態感測與最佳化軌跡生成

Abstract

本論文嘗試改良舊有的袋鼠機器人，藉由參考實際的袋鼠身體結構，在原有的尾巴架構上增加一個滑動自由度，以此來調整附件(尾巴)對系統的影響，使得系統總質心位置移動和相對轉動慣量大小變化。在模型方面進一步的由單純原型腳被動模型(RSLIP)延伸為帶有多質點且偏心的複雜系統，以納入尾巴和腳在模擬中的表現；系統的總質心在運動的過程中產生相對的位移，使動態模型具有相當高的非線性度，難以進行變數分離分析對系統的影響，因此採用間接最佳化成本函數的方式來得到軌跡，根據在地段和載空段的差別分別使用黃金比例搜尋法和粒子演算法進行最佳化。即使在模型複雜化後，在實驗下仍可看出實際系統與模型之間的差異，在此篇論文中進一步介紹以觸地調整閉迴路步態和非單一穩定點軌跡調整，而為了配合該方法的實行，在估測系統方面提出藉由預測軌跡來減少時間延遲的高斯濾波器，和針對定角速度模型的混合卡曼濾波器架構。除了上述所提到針對舊有系統的改良，本篇論文額外以光流場演算法為基礎提出視覺姿態估測器，提供未來研究中有機會以類似動物視覺的方式進行動態回饋。

The project attempts to improve the old hopping robot, by examining the actual physical structure of kangaroo. The robot adds an additional degree of freedom which is a sliding mass attached to the tail, so that the robot has the capability of adjusting the center of mass and the inertia relative to the overall system. On the aspect of simulation model, the project extends the simple passive model (RSLIP) to an eccentric and multiple parts system, including body, tail and feet. Due to the displacement of overall system’s center of mass, the model has very high nonlinearity. The highly coupled variables are difficult to analyze separately, thus the project using an indirect way to get the trajectory by optimizing the cost function. Though the simulation runs a relative complex model, the difference between the real system and the model can be seen under the experiment. Therefore, the project further proposes two methods to modify the trajectory, including a touchdown based trajectory adjustment and a dynamic trajectory selection. In order to meet the requirement of the method proposed, a predicted trajectory Gaussian filter was implemented to reduce the white noise of the sensor, and a hybrid Kalman filter architectures with fixed angular velocity model for more precise estimation. In addition to the improvement of the robot mentioned above, the thesis also proposes a posture estimator which is based on optical flow algorithm. Though this estimator has not been used in the project, it provides a scenario that robot can use dynamic visual feedback for control in the future.