Purpose The application of industrial robots in modern production is becoming increasingly widespread. In the context of flexible production lines, quickly and accurately identifying and grasping specified workpieces is particularly important. This study aims to propose a grasping scheme that combines traditional methods with deep learning to improve grasping accuracy and efficiency. Design/methodology/approach First, a dataset generation method is proposed, which constructs a point cloud dataset close to the real scene without the need for extensive data collection. Then, the 3D object detection algorithm PointPillars is improved based on the features of the scene point cloud, allowing for the analysis of part poses to achieve grasping. Finally, a grasp detection strategy is proposed to match the optimal grasp pose. Findings Experimental results show that the proposed method can quickly and easily construct high-quality datasets, significantly reducing the time required for preliminary preparation. Additionally, it can effectively grasp specified workpieces, significantly improving grasping accuracy and reducing computation time. Originality/value The main contribution of this paper is the integration of a novel dataset generation method, improvements to the PointPillars algorithm for 3D object detection and the development of an optimal grasp detection strategy. These advancements enable the grasping system to handle real-world scenarios efficiently and accurately, demonstrating significant improvements over traditional methods.
Read full abstract