PhD - End-2-End Trainable System for Autonomous Driving with Introspection

Current automated driving systems are structured in a hybrid manner, meaning that classical model-driven approaches are combined with data-driven AI approaches. So far all modules are developed or trained independently based on module-related key performance indicators. As a second step system-related evaluation is conducted in software-in-the-loop tests and closed-loop in the vehicle. The main drawbacks are on the one hand, that the independent optimization of modules offers no guarantee of a global system-related optimum and on the other hand, the manual experience-based analysis of the resulting system to determine the performance-related weak points in the architecture. The slow system design process is also a disadvantage.

• The goal of this PhD thesis is to research the system aspects of a ground breaking innovation in system design: End-2-End trainable systems.
• You will edit following research questions: How to efficiently train a system that is composed of AI sub-modules (efficient loss propagation)? How to train a hybrid system? And you will also research safety-related measures that allow introspection of the system.
• In your PhD thesis you will develop novel machine ...