Application of neural networks to the collision avoidance problem in 2D based on the TensorFlow library

Abstract

The objective of this paper is the learning and initiation into the world of neural networks using the Google tool, TensorFlow. In order to do this, we consider a series of algorithms whose purpose is the control of a drone which can move in a specific environment, avoiding static and mobile obstacle while, at the same time, guaranteeing a safe navigation. This tool is the main different with respect to older research in this field. Furthermore, we look into the structure and the diverse tools that this platform offers with the intention of discovering the areas in which TensorFlow can be useful. Therefore, the organization of this paper is structured as follows: In the first place, we offer an introduction that covers Neural Networks that are so important nowadays in the wide range of application available. We also explain what they are based on and how the information is used. Next, TensorFlow structure is briefly introduced, explaining also how it works and some of the basics tools provided by it. After that, the setting in which we are currently working is illustrated in three different steps. First, a data set is created, then the TensorFlow algorithms are implemented for the different scenarios and finally the "learning" obtain by the neural networks are analysed. Lastly in our conclusions we offer two significant points: first, we demonstrate the findings of the different neural networks in the simulator provided and, second, the conclusions that we have reached in this paper and the future line of researches that this study put forth.