The classification and grasping of randomly placed objects where only a limited number of training images are available, remains a challenging problem. Approaches such as data synthesis have been used to synthetically create larger training data sets from a small set of training data and can be used to improve performance. This paper examines how limited product images for ‘off the shelf’ items can be used to generate a synthetic data set that is used to train a that allows classification of the item, segmentation and grasping. Experiments investigating the effects of data synthesis are presented and the subsequent trained network implemented in a robotic system to perform grasping of objects. © Springer International Publishing AG, part of Springer Nature 2018.