Advancing Digital Twins: Accelerating and Enhancing Workflows in Bin Picking Applications.

In recent decades the fields of computer vision and industrial automation have been undergoing unprecedented evolution. As a result, applications that were merely theoretical hypotheses are now reality. Thanks to better controllers, sensors and computational power, robots are able to perform increasingly complex tasks without human intervention or supervision. One such application is ”random bin picking,” that is the recognition and subsequent manipulation of objects whose placement is unknown in a bin. To obtain the data of the object to be searched, expensive scanners or 2D cameras are used to generate datasets on real samples of it. In recent years, to reduce the time and cost of generating datasets needed for model training and tuning of search algorithms, more and more is being invested in the generation of virtual datasets and the creation of digital twins. In particular the necessary point clouds or 2D views necessary for those algorithms, will be generated starting from a 3D model of the object. However, the 3D models provided, often contains geometry that is not useful for the dataset generation. Such geometry should be removed to speed up the process and avoid training errors. Real cameras also contain imperfections in their lenses and alignment that must be simulated to generate valid datasets. This work therefore aims to achieve two goals: 1. To provide a solution to simplify generic 3D models by removing invisible geometry from the outside; 2. To provide a solution for simulating the radial distortion of 2D camera lenses. This research work was carried out in collaboration with the company Euclid Labs of Nervesa della Battaglia, which over the years has specialized in offering solutions for bin picking.