Skip to Main content Skip to Navigation

Apprentissage par renforcement du contrôle d’un véhicule autonome à partir de la vision

Abstract : In this thesis, we address the challenges of autonomous driving in an urban environment using end-to-end deep reinforcement learning algorithms. Reinforcement learning (RL) is one of the three paradigms of machine learning. It distinguishes itself from supervised learning by the fact that agents learn by trial and error from a reward signal and not by supervision with input-label pairs. In reinforcement learning, we explicitly seek to optimize sequences of actions in order to maximize long-term behavior. The major advantage of RL is that the agent learns the behavior to be followed by exploring and interacting with his environment : we therefore do not need to explicitly indicate the actions to be taken. First, we proposed a new reinforcement algorithm, Rainbow-IQN Ape-X, by combining three major articles in the field of Value-based Reinforcement Learning. This algorithm achieves state-of-the-art performance on the Atari benchmark. Using this distributed reinforcement algorithm, we introduced a new method coined implicit affordances, which allows to train by reinforcement neural networks with more parameters and larger inputs than previous works in DRL. This technique allowed us to demonstrate for the first time a reinforcement algorithm capable of driving in a complex simulator including pedestrians, vehicles and especially traffic lights. Finally, we used all of our previous contributions to perform real data reinforcement learning for urban driving. The fundamental idea of our approach is to use a simulator based on real images to successfully train agents capable of generalizing to real data.
Document type :
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Friday, September 17, 2021 - 12:19:10 PM
Last modification on : Wednesday, November 17, 2021 - 12:31:12 PM
Long-term archiving on: : Saturday, December 18, 2021 - 6:40:14 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03347567, version 1


Marin Toromanoff. Apprentissage par renforcement du contrôle d’un véhicule autonome à partir de la vision. Robotique [cs.RO]. Université Paris sciences et lettres, 2021. Français. ⟨NNT : 2021UPSLM020⟩. ⟨tel-03347567⟩



Record views


Files downloads