以環境RGBD資訊進行輪腳複合機器人之地形判定導航與步態選擇

Abstract

根據外在環境變化調控運動模式，一直是足類機器人持續研究的方向。本研究致力於使用即時色彩及深度影像回授系統，讓具有特殊輪腳變換機構之四足機器人TurboQuad-V，在既有的中樞模式控制架構下，根據不同地表性質與幾何資訊，進行自主步態與方向調控。 透過影像處理演算法，可將即時影像轉換為動態切換參考依據。色彩影像的部分，建構卷積神經網路分類架構，並自行收集地形資料庫進行訓練，用以區分不同地貌種類；深度影像的部分，則透過點雲回推地表幾何分布，計算機器人行進路徑，及路徑上幾何分布情況，進而獲取轉向資訊及地表崎嶇程度資訊。 本研究亦開發二維單足輪腳模擬，考慮複雜地表幾何及打滑因素，以實際步態在簡化之地表模型進行動態模擬，並分析步態之能量效率、越障能力、運動穩定度，進而歸納各步態最佳切換操作點。 經動態實驗證實，綜合影像演算法，搭配模擬與實驗共同驗證之步態切換法則，可以讓四足機器人在多元地表環境中，達到即時自主步態與方向調控。

Adjusting dynamic motions according to the environment has been an important research topic in recent years. The main contribution of this thesis is the development of autonomous gait and direction switching on a leg-wheel transformable quadruped robot, TurboQuad-V, by using real-time color and depth information of the terrain combined with the original bio-inspired CPG structure. Color image is used to classify different types of landscape by using Convolution Neural Network model, which is trained with the database collected in this research. Depth image is transferred to elevation grids to represent the geometry distribution of the terrain, and the optimized path as well as the height distributions along the route can be calculated. This research also focuses on the development of the dynamic simulation using single-leg-wheel model. The simulation runs different gaits on various simplified terrains while considering slip effect and contact geometry. With the analysis of indexes such as power efficiency, the ability to overcome obstacles and the height variation during motions, suitable operating points of each gait according to different kinds of terrain can be concluded. With the integration of information from the image processing and the switching policy of each gait, TurboQuad-V is proved to have the ability to perform autonomous gait switching and basic obstacle avoidance by the experiments conducted on multiple environment.