Enhanced Grey Wolf Optimizer based Multiple Object Grasping Poses for Home Service Robot

Abstract

The thesis proposes an Enhanced Grey Wolf Optimizer (EGWO) to learn multiple grasping poses of unknown objects for a home service robot For accomplishing this task a 3D model of an object should be established at first The depth information obtained from Kinect is converted to 3D points in every frame and is matched by Iterative Closest Point (ICP) to track the pose of Kinect The matched points are then integrated to a volumetric surface and the result is presented by ray casting However the result of 3D object model is too complex to calculate grasping pose So that a simplified method is proposed in this thesis The original 3D object model is transferred to a triangle mesh firstly and the triangle vertices are classified by three-stage nearest neighbor algorithm to find surfaces of the object Therefore the simplified surfaces can be constructed by the least square method and the object can be described by these simplified surfaces After established a 3D object model this thesis proposes Enhanced Grey Wolf Optimizer (EGWO) to learn multiple grasping poses Due to the original Grey Wolf Optimizer (GWO) is highly corresponding with the origin of searching space the performance will be influenced by the place of global optimal The proposed EGWO solves the problem by eliminating the influence of the origin In addition it adds a position error term to maintain the good exploration and exploitation The position error term is calculated by the difference between current and previous positions If the difference is large the algorithm forces more on exploration Contrarily the algorithm forces more on exploitation Both the simulations and robotic experiments demonstrate that EGWO provides much better performance than GWO on learning multiple grasping poses and makes the home service robot successfully grasp unknown objects

Date of Award	2016 Aug 18
Original language	English
Supervisor	Tzuu-Hseng S. Li (Supervisor)