Master Thesis - Provably Safe Adaptive Reinforcement Learning for Autonomous Driving

23.04.2026, Abschlussarbeiten, Bachelor- und Masterarbeiten

Safe reinforcement learning (RL) has shown strong potential for decision making in complex and uncertain environments. However, for safety-critical systems, achieving high performance is not sufficient: the learned policy must also remain provably safe under changing environmental conditions. To address this challenge, this thesis focuses on extending a framework that combines learning-based action masking with formally verified fail-safe maneuvers. The method has already been tested in abstract benchmark tasks, to further demonstrate its applicability and effectiveness in more challenging scenarios, we will apply it to autonomous driving scenarios.

The student will build on our existing framework and implement longitudinal and lateral action masking networks to prevent driving off-road and guarantee safe-distance with the preceding vehicles. The resulting method will be evaluated under various scenarios, including different road shapes and traffic conditions, and the adaptability will be assessed.

Details can be found in the attached file.

Kontakt: shuaiyi.li@tum.de

Mehr Information