Master Thesis Combining Imitation & Reinforcement Learning to Solve Automated Driving

Imitation Learning (IL) and Reinforcement Learning (RL) each come with distinct strengths and weaknesses. IL is typically sample-efficient and straightforward to implement, but it requires large amounts of expert data and often suffers from distributional shift. In contrast, RL does not rely on expert demonstrations and can learn robust policies through interaction, but it faces challenges such as unstable training dynamics and the difficulty of designing appropriate reward functions. IL has since long been applied to the problem of autonomus driving. Since recently, also RL is getting more traction, among others due to the availability of extremely fast simulators and large-scale compute. Goal of this thesis is to investigate the emergent topic of combining both approaches.

• During your thesis, you will conduct in-depth literature research and selectively implement existing approaches in the field of reinforcement and imitation learning.
• Furthermore, you will extend established methods and ideas to create innovative solutions that go beyond mere reimplementation.
• In addition, you will explore the generation of multi-step models using reinforcement learning to optimize algorithms for use in vehicles.
• You will combine reinforcement learning, particularly single-step models, with imitation learning to generate precise trajectories for autonomous driving functions.
• Lastly, you will carefully document your research findings and prepare them for scientific publication.