High-level Decision Making under Safety Constraints for Autonomous Vehicles

Future transportation systems serve as a crucial foundations for enhancing human productivity. Among these, automated driving has emerged as a vital technology with the potential to improve the comfort and efficiency of road traffic while freeing human hands from driving tasks.

It is anticipated that automated driving systems will coexist with human drivers on the road for decades before achieving full automation. As such, a key research goal is to develop autonomous vehicles that closely mimic human driving behavior, enabling passengers, other human drivers, and traffic participants to better understand and cooperate with these vehicles. Furthermore, ensuring provable safety is essential for the widespread acceptance of automated driving systems.

To develop driving behavior capable of handling generic traffic scenarios, existing approaches often frame the behavior planning problem as a sequential decision-making process aimed at maximizing expected future rewards. These methods frequently suffer from unrealistic reward functions and lack definitive proof of human-like behavior. To address these shortcomings, machine-learning techniques have been employed to derive driving policies from recorded human driving trajectories in real traffic. ... mehrHowever, some of these approaches face runtime challenges, while others struggle to ensure compliance with traffic rules and safety regulations.

In this work, I introduce a high-level decision-making framework for autonomous vehicles, focusing on safety across diverse traffic situations and adherence to traffic rules. This method, enhancing an existing safety concept, takes into account factors like road types and occlusions, while also relaxing safety requirements to achieve more natural driving behavior. It involves generating safe action options, simulating future traffic scenarios to assess outcomes, and using machine learning to analyze human driving data for decision-making. This results in actions that mirror human decision processes in complex driving environments.

My approach is evaluated through simulations of various scenarios, including parallel lanes and intersecting lanes. Simulation evaluations demonstrated that the learned policy outperformed the rule-based baseline approaches, producing more human-like behavior while balancing driving efficiency, comfort, perceived safety, and politeness.

Finally, the key part of the proposed approach has been successfully deployed on real experimental vehicles. Demonstrations have been conducted on a test track, as well as in regular on-road experiments.