A generic visual perception domain randomisation framework for Gazebo

Abstract

The impressive results of applying deep neural networks in tasks such as object recognition, language translation, and solving digital games are largely attributed to the availability of massive amounts of high quality labelled data. However, despite numerous promising steps in incorporating these techniques in robotic tasks, the cost of gathering data with a real robot has halted the proliferation of deep learning in robotics. In this work, a plugin for the Gazebo simulator is presented, which allows rapid generation of synthetic data. By introducing variations in simulator-specific or irrelevant aspects of a task, one can train a model which exhibits some degree of robustness against those aspects, and ultimately narrow the reality gap between simulated and real-world data. To show a use-case of the developed software, we build a new dataset for detection and localisation of three object classes: box, cylinder and sphere. Our results in the object detection and localisation task demonstrate that with small datasets generated only in simulation, one can achieve comparable performance to that achieved when training on real-world images.