To achieve full autonomy in self-driving cars and other autonomous vehicles, different approaches are being researched on how autonomous driving systems should be implemented. As an alternative to a traditional modular approach, an autonomous driving system can be implemented in an end-to-end manner using a deep neural network that takes sensory information as an input and provides control signals as an output. Such systems are usually trained in a simulation environment. In this paper, the author proposes how to train an end-to-end autonomous driving system prototype using the Unity platform and Unity ML-Agents toolkit. After successfully implementing a prototype of such a system, the author provides recommendations on how to train end-to-end autonomous driving systems using Unity as a simulation platform. The experiments conducted in this thesis suggest that successful training can be achieved by combining reinforcement learning with imitation learning, using diverse initial locations for the vehicle model during the training, and providing recorded demonstrations that capture diverse situations of correct driving.
Ištisinių autonominių vairavimo sistemų mokymas naudojant „Unity“ kaip simuliacijų platformą ; Training end-to-end autonomous driving systems using unity as simulation platform
2020-06-10
Theses
Electronic Resource
Lithuanian , English
DDC: | 629 |
Dúšok jedu : takmer generačná platforma
GWLB - Gottfried Wilhelm Leibniz Bibliothek | 1997
|DataCite | 1970
|TIBKAT | 1900
|