Research On Robot Autonomous Grasping Technology Based On 3D Point Cloud Data

With the rapid development of society,human beings have an increasingly urgent need for robots.However,the lack of scene interaction ability seriously limits the popularity of robots.Grasping of unknown objects with neither RGB data nor object models in advance is very important for robots that work in an unfamiliar environment.The key to autonomous grasping of robots lies not only in the recognition of object types,but also in the good grasping of object's shapes(such as the composition of shape primitives).Aiming at the unstructured grasping environment in real life,an autonomous grasping method based on three-dimensional point cloud data is proposed.The simplification of complex objects helps to describe the shapes of complex objects,thus providing ideas for the selection of grasping strategies and improving the accuracy of autonomous grasping of robots.Firstly,3D camera is used to obtain local point cloud data of the target object.Secondly,the target object point cloud data is simplified,and proposed the 3D point cloud simplification algorithm based on the basic shape.The 3D data points of the grasped object are segmented into the main body and the branch part based on the mesh segmentation using feature point and core extraction(MFC)algorithm,and the parts are fitted to a sphere,an ellipsoid,a cylinder or a parallelepiped according to an optimal fitting algorithm,so as to simplify the complex object.Then,according to the simplified shape,the grasping posture of the robot is restrained.This paper propose a grasp strategy based on the basic shape and suggest to grasp the main parts of the object.And select grasping candidates based on shape constraint.Finally,set up grasping data-set,and the main part of the object is used for grasping training using deep learning,and build a new grasp description that incorporates grasp heightmaps and angle graph of object surface to measure the grasp similarity.Finally,the method proposed realizes the grasping operation with an accuracy of93.3% on the Baxter robot.In our benchmark,compared with current grasp quality classification methods,this paper need fewer training sets to achieve higher grasp success accuracy.And the experimental results show that the method can be applied to unknown irregular objects of different shapes and postures with strong robustness.