Synthetic Data Generation for Robotic Order Picking

Abstract

Advances in robotics, especially in computer vision, have led to the increasing use of robots in order picking. Deep Learning methods using CNN for computer vision purposes have shown good object detection and localization results. However, training neural networks requires a large amount of domain-specific labelled data. In this work, we generated synthetic data and converted it to the appropriate format to be fed to neural network. For this purpose, randomized camera angles, backgrounds, and object configuration are used for data augmentation. A generalized and balanced dataset is ensured by varying these parameters based on the properties of natural objects.