Sorting method for guiding mechanical arm to grab materials with different poses based on ConvPoint model

Abstract

The invention discloses a sorting method for guiding a mechanical arm to grab materials with different poses based on a ConvPoint model, three-dimensional point cloud data of the materials is obtained through a 3D structured light camera, the three-dimensional point cloud data is combined with the ConvPoint model for material sorting, the method is not affected by the initial poses and appearances of the materials, various grabbing targets are achieved, and the system robustness is higher. Compared with an existing deep learning method, the method has the advantages that grabbing sampling is improved through PCA, the material grabbing accuracy is obviously improved, in addition, the point cloud in the network input clamp is evaluated instead of the point cloud of the whole object, the learning and reasoning time is shortened, and the material sorting efficiency is greatly improved. Besides, in the ConvPoint model training stage, the point cloud data of different types of complex structure materials are collected in real time to serve as a training set, the method does not depend on a data set of simple objects of an online open source, complex objects of different poses are grabbed, and the intelligent level and the automation level of a sorting system are improved.

Description

technical field [0001] The invention relates to a material sorting method based on a ConvPoint model, in particular to a sorting method based on a ConvPoint model to guide a mechanical arm to grab materials in different poses. Background technique [0002] At present, with the popularization of consumer-grade visual equipment, it has become a hot topic to use the camera to guide the robotic arm to complete the grasping. Among them, generating the 6D grasping pose (i.e., 3D position and 3D orientation) of the actuator is the key for any robotic arm to accomplish grasping. However, most of the current research is based on 2D images or 2.5D depth images. Compared with 3D point clouds, images and depth images lack three-dimensional spatial geometric information, data features are not obvious or feature points are relatively small, and complex manual features are sometimes required to process data. The grasping of the robotic arm is very dependent on the three-dimensional space...

AI Technical Summary

Problems solved by technology

However, it is difficult to estimate the normal of the target surface when the noise is relatively large or the point cloud is sparse, and it is impossible to achieve accurate high-quality sampling. For example, it is difficult to generate grasping candidates along the edge structure of a cup. Therefore, there is an urgent need for a method that can Effectively improve the means of grabbing pose sampling

Images