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

The current state of the art in autonomous driving (AD), represented by Vision-Language-Action (VLA) models, has enabled vehicles to link perception with action in a more data-efficient way than classical End-to-End systems. Still, we are reaching the limits of VLAs: short-sightedness due to a lack of long-term prediction capabilities, no physical consistency, and limited coverage of long-tail scenarios. Therefore, another paradigm shift is required.

The aim of this research is to investigate the next generation of AD systems by moving from reactive VLA to predictive, agentic world models. The core thesis is that "imagination before action" is essential for safe and scalable autonomy.

• As part of our team, you will develop a framework for integrating VLA, world models, and agentic planners into a cohesive driving system.
• Furthermore, you will demonstrate improved performance in long-term prediction, edge-case handling, and verifiable safety through proactive simulation.
• You will drive disruptive change in how data is prepared and large systems are trained more efficiently by focusing on unexpected scenarios.
• Last but not least, you will contribute to the fundamental shift in the AD field from reactive control to imagination-based, agentic action, paving the way for more capable and trustworthy autonomous systems.

Function: Research
Experience Level: Entry Level