École Nationale Supérieure de
Techniques Avancées

Abstract

This report describes the work executed during an engineering internship at AKKA autonomous vehicles group. The project conducted during this time consisted in implementing a test platform for end-to-end driving policies for autonomous vehicles, in the form of a software architecture and also a physical prototype. This kind of approach has been gaining relevance in the last years due to the advances in the machine learning field, making possible a direct association between sensor data and driving commands through deep learning techniques. The vehicle’s architecture was developed in ROS, with a simulation on Gazebo and interface to the driving agent achieved through the OpenAI Gym toolkit. System behavior was validated in the conducted tests, and the result is a modular and flexible design, where developing driving algorithms is independent of low level vehicle dynamics and the nature of the environment. Though preliminary tests with a Q-learning algorithm did not give the expected results, the studied approach has potential to improve after addressing the difficulties exposed in this text.