End-to-end control on F1/10 Autonomous Car using Neural Network
End-to-end řízení F1/10 autonomního auta s využitím neuronových sítí
Authors
Supervisors
Reviewers
Editors
Other contributors
Journal Title
Journal ISSN
Volume Title
Publisher
České vysoké učení technické v Praze
Czech Technical University in Prague
Czech Technical University in Prague
Date
Abstract
Táto práca sa zameriava na autonómne riadenie modelu vozidla F1TENTH v prostredí pretekov s jedným agentom. Po preskúmaní už existujúcich riešení boli vybrané dve metódy. TD3, algoritmus ktorý nepoužíva model, a Dreamer, algoritmus ktorý sa spolu s reakciami snaží odhadnúť aj model okolia. Opísané sú prostredia pre simuláciu a reálne scenáre experimentov. Metódy sú opísané spolu s teóriou potrebnou na ich pochopenie. Uvádzajú sa trate použité ako tréningové a testovacie scenáre pre experimenty. Sú uvedené časy výsledkov s trajektóriami zo simulácie. Výsledky sú zdokumentované a diskutované. Oba agenti úspešne zvládli prejsť kolo s lepším časom ako Follow the Gap algoritmus ktorý bol použitý pre porovnanie. Na záver sa navrhuje niekoľko vylepšení simulačného modelu, ako aj algoritmov bez modelu a algoritmov založených na modeli.
This thesis focuses on autonomous end-to-end control of an F1TENTH model car in a single-agent racing environment. After reviewing already existing solutions, two methods were selected: a model-free algorithm, TD3, and a model-based algorithm, Dreamer. The environments for simulation and real-life scenarios are described. The methods are described along with the theory needed to understand them. The tracks used as training and testing scenarios are presented. The result times are shown with the trajectories from the simulation. The results are documented and discussed. Both agents trained on the CIIRC track managed to safely complete a lap with a better time than the Follow the Gap algorithm provided as a baseline. Finally, a few improvements to the simulation model, as well as to both model-free and model-based algorithms, are proposed.
This thesis focuses on autonomous end-to-end control of an F1TENTH model car in a single-agent racing environment. After reviewing already existing solutions, two methods were selected: a model-free algorithm, TD3, and a model-based algorithm, Dreamer. The environments for simulation and real-life scenarios are described. The methods are described along with the theory needed to understand them. The tracks used as training and testing scenarios are presented. The result times are shown with the trajectories from the simulation. The results are documented and discussed. Both agents trained on the CIIRC track managed to safely complete a lap with a better time than the Follow the Gap algorithm provided as a baseline. Finally, a few improvements to the simulation model, as well as to both model-free and model-based algorithms, are proposed.