Deep Reinforcement Learning for Autonomous Driving: Human-Informed, Ethical, and Transferable Agents

Dianzhao Li

AutorIn

Dianzhao Li

Titel

Deep Reinforcement Learning for Autonomous Driving: Human-Informed, Ethical, and Transferable Agents

Zitierfähige Url:

https://nbn-resolving.org/urn:nbn:de:bsz:14-qucosa2-1048360

Erstveröffentlichung

2026

Datum der Einreichung

01.10.2025

Datum der Verteidigung

19.05.2026

Abstract (EN)

This thesis investigates how deep reinforcement learning (DRL) can be used to improve planning and control in autonomous driving (AD) systems. While DRL has shown promise in learning complex driving behaviors, practical deployment faces three persistent challenges: the sensitivity of agent behavior to reward function design, the lack of mechanisms to account for ethical responsibility, and difficulties in transferring trained agents from simulation to real-world platforms. To address these challenges, we present a series of contributions across simulation-based and real-world experiments. First, we introduce a two-stage DRL framework that leverages human driving data to enhance car-following behavior, enabling agents to learn from demonstrations and reduce reliance on handcrafted reward functions. Second, we propose a hierarchical Safe RL framework that integrates ethical risk modeling and prioritized learning from safety-critical scenarios, demonstrating that AV agents can reduce ethical risks while maintaining performance. Third, we provide a systematic survey of small-scale autonomous car platforms, classifying approaches into modular and end-to-end categories and identifying trends for future research. Fourth, we design a vision-based DRL agent capable of simultaneous car-following and lane-keeping, which is validated through both simulation and deployment on small-scale platforms. Finally, we extend this approach to the more complex overtaking task using Long Short-Term Memory (LSTM)-based DRL, showing that the framework generalizes across platforms and remains robust under simulation-to-reality (Sim2Real) gaps. The findings of this thesis demonstrate that DRL-based AD systems can be made more adaptable, ethically aware, and practically transferable. By combining human-informed learning, safety-aware design, and real-world validation, this work contributes to the development of autonomous vehicles that are not only technically capable but also aligned with human values and societal expectations.

Freie Schlagwörter (EN)

Autonomous Driving, Reinforcement Learning

Klassifikation (DDC)

380

006

Klassifikation (RVK)

ZO 4660

ST 300

GutachterIn