The request for small-size batches and customized products is pushing the need for flexible production lines capable of switching between products flawlessly and with no human intervention. The use of robotic sorters is widespread in different industries and, in most of them, there is a computer vision (CV) system with the objective to perform an object detection task. However, the implementation of Artificial Intelligence (AI) is limited to some particular usage. The proposed idea is to join Robotics with Deep Learning (DL), with the aim of creating an automatic flexible sorting system, capable of handling a high variability in terms of shape, position, and class, avoiding stopping production when a new item is introduced. An automatized solution is developed, able to detect new products on the conveyor and to directly generate online labeled real images to be used to re-train the deep learning network. Starting from this point, synthetic images can also be generated. In the experimental campaign, various training trials are carried out to define the best balance between real and synthetic images, taking into consideration performance and time efficiency.
Deep learning-based robotic sorter for flexible production
Zaki A. M. A.;Carnevale M.;Giberti H.
2022-01-01
Abstract
The request for small-size batches and customized products is pushing the need for flexible production lines capable of switching between products flawlessly and with no human intervention. The use of robotic sorters is widespread in different industries and, in most of them, there is a computer vision (CV) system with the objective to perform an object detection task. However, the implementation of Artificial Intelligence (AI) is limited to some particular usage. The proposed idea is to join Robotics with Deep Learning (DL), with the aim of creating an automatic flexible sorting system, capable of handling a high variability in terms of shape, position, and class, avoiding stopping production when a new item is introduced. An automatized solution is developed, able to detect new products on the conveyor and to directly generate online labeled real images to be used to re-train the deep learning network. Starting from this point, synthetic images can also be generated. In the experimental campaign, various training trials are carried out to define the best balance between real and synthetic images, taking into consideration performance and time efficiency.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
