資料庫智慧型抓握

Abstract

在不久的將來，將會有越來越多的機器人代替人類完成各式各樣的任務，多指機器手掌將會是其中的一環。多指機器手掌的自由度比起機器爪多了許多，如何在有限的時間裡，有效地抓取各式各樣的物體變得十分重要。因此，本論文旨在探討如何透過建立資料庫的方式，辨識出三維點雲中的物體以及其姿態，並搜尋該姿態下機器手掌對應的抓握姿勢，最後利用兩顆快速擴展隨機樹連接法，在手臂的六軸空間中尋找避開障礙物的路徑。對於不在資料庫裡的物體，也可以根據邊界盒以及力旋量空間中的品質評估，透過模擬器來搜尋最終的抓握姿勢。

In the near future, more and more robots will substitute human to serve tasks, multi-fingered robot hand will be one of the robots. Since the degrees of freedom of a dexterous hand are more complex than a simple gripper, how to grasp various types of objects efficiently in a short time has become more and more important. This thesis focuses mainly on how to recognize the objects and their poses in three-dimensioned point cloud based on construction of the database. Then the grasp posture of the robot hand will be searched in the database in terms of the object pose. Finally, the rapidly-exploring random trees (RRT) – connect is used to search for the collision-free path in the joint space of the robot arm. For those objects not in our database, we can also use the simulator to find the final grasp posture according to wrench space and quality measure.