Robustness of reinforcement learning based autonomous driving technologies

Fabian Hart

authors00000

Fabian Hart

title

Robustness of reinforcement learning based autonomous driving technologies

Please use the following URL when quoting:

https://nbn-resolving.org/urn:nbn:de:bsz:14-qucosa2-961296

publication_date

2025

Date of submission

12.01.2024

Date of defense

10.12.2024

Abstract (EN)

Autonomous driving technologies offer the potential to substantially improve safety, mobility, and sustainability in the field of transportation. However, the complex and unpredictable nature of real-world driving scenarios demands robust models capable of making safe decisions across diverse and unseen situations. With the rise of machine learning techniques in the recent decade, reinforcement learning is an increasingly used method to tackle the challenging nature of au- tonomous driving tasks. This thesis contributes to the ongoing advancements in reinforcement learning based autonomous driving technologies by offering valuable insights into the design of training environments with a particular focus on safety robustness, which refers to the model’s capability to guarantee safety in diverse scenarios, which may not have been encountered in training. This thesis provides experiments for vehicle-following and obstacle avoidance tasks, which demonstrate that reinforcement learning based models struggle to learn effectively from natural driving data or naturally inspired training environments. In both cases, safety-critical situations are rare, which leads to poor performance in extreme situations, also known as the curse of rarity. To combat this issue, this thesis discusses synthetic training approaches for different reinforcement learning based applications of autonomous driving that increase safety- critical situations in training. Validating the trained models in various scenarios, qualitatively different from the training data, and under extreme conditions, demonstrates their capability to be robust. Furthermore, this thesis suggests different approaches in the design of the objec- tive functions, which should be optimized, and the design of environment observations, from which the model learns, in order to improve safety robustness, for example, by incorporating time-to-collision metrics for vehicle-following and closest-point-of-approach metrics for obstacle avoidance.

Keywords (EN)

reinforcement learning, autonomous driving, robustness

Classification (DDC)

380

Classification (RVK)

ZO 4660

Examiner