Blayze Millward Undergraduate Dissertation 2017/18

Using Simulated Environments to Learn Real World Navigation

Supervised by J.Barker

Abstract

Autonomous navigation of robotic, driverless agents acting in the real world is a rapidly expanding area of research and business interest. Almost all modern approaches to autonomous navigation rely on the use of artificial neural networks – function estimators inspired by the connections of neurons found within biological brains. A major downside to the use of artificial neural networks, and machine learning approaches in general, is their reliance on training data, i.e. large corpuses of example data, in order to successfully estimate functions. These corpuses can be hard, costly and sometimes infeasible to collect. This report will look at the required techniques and technologies to asses the suitability of simulated environments as a means to generate this training data instead. This report analyses historic and current machine-learning based approaches to solving the problem of machine navigation as well as evaluating current available technology with the aim of outlining a project plan that will test the use of a 3D virtual environment to train a neural network to solve real world navigation problems.