Autonomní řízení vozidel

Abstract

Autonomous driving has the potential to increase the efficiency of the transport system by completely eliminating the human factor. Autonomous driving reduces the ecological footprint through more environmentally friendly behavior in the form of lower emissions, in addition to making transport more accessible, especially for people who cannot drive alone due to health restrictions. The aim of the work is to create a car simulator for testing reinforcement learning algorithms with an application in autonomous control. Specifically, in my work I try to use the Q-Learning algorithm in combination with a deep neural Q network to learn to control the vehicle from his own experience gained through trial and error. Two experiments were performed. The first tests the vehicle to teach the vehicle a racing driving style and the second, on the contrary, a smooth and safe driving style. The control learning problem was divided into two phases, each phase using its own instance of a convolutional neural network and being trained separately in the chosen order. All training was carried out in a procedurally generated environment with the vehicle inserted into it. A vehicle with user-definable physical properties and parameters allows the algorithm to be tested with different types of vehicles. The input of the models are images taken by a front camera, which captures the immediate surroundings of the vehicle and simulates the driver's view. The output of the work is a system for simulating driving of various complexity, which is able to share information about the situation around the vehicle and at the same time allow to control this vehicle using a reinforcement learning algorithm called Q-Learning.