Visual Data Simulation for Deep Learning in Robot Manipulation Tasks

This paper introduces the usage of simulated images for training convolutional neural networks for object recognition and localization in the task of random bin picking. For machine learning applications, a limited amount of real world image data that can be captured and labeled for training and testing purposes is a big issue. In this paper, we focus on the use of realistic simulation of image data for training convolutional neural networks to be able to estimate the pose of an object. We can systematically generate varying camera viewpoint datasets with a various pose of an object and lighting conditions. After successful training and testing the neural network, we compare the performance of network trained on simulated images and images from a real camera capturing the physical object. The usage of the simulated data can speed up the complex and time-consuming task of gathering training data as well as increase robustness of object recognition by generating a bigger amount of data.