Bachelor/Semester/Master thesis: Risk-aware scenario analysis based on LLMs in autonomous driving

14.11.2024, Diplomarbeiten, Bachelor- und Masterarbeiten

The primary objective of this project is to develop a framework that uses LLMs for risk-aware analysis of generated driving scenarios.

Background

The autonomy of vehicles has advanced rapidly in recent years, reaching a level where human intervention is barely or not at all required in certain controlled environments. Leading the way are vehicle manufacturers such as Mercedes and BMW, which offer autonomous road vehicles at Level 3, depending on the system's design. This progress is largely based on the development and validation of highly reliable driving functions. Ensuring the safety and reliability of these driving functions requires extensive testing in diverse and challenging scenarios.

Scenario-based testing has emerged as a cost-effective method for simulating realistic driving conditions in this context. It allows for assessing how well a driving function meets the required demands. However, a critical challenge lies in determining whether these scenarios are sufficiently safety-critical—i.e., whether they cover the full spectrum of necessary safety assurances. This is essential to ensure that driving functions can handle rare or even unthinkable but dangerous driving situations, thereby enhancing their robustness and reliability.

Current Gaps in Scenario-Based Testing

· Limited Assessment of Safety-Criticality: Existing methods often fail to accurately determine how safety-critical a scenario is, potentially overlooking important edge cases.

· Lack of Feedback Mechanisms: Once a scenario has been evaluated, there are limited opportunities to refine it based on its safety-critical assessment.

· Manual Analysis Constraints: Evaluating scenarios often involves considerable manual effort, which is both time-consuming and prone to subjectivity.

To address these challenges, Large Language Models (LLMs), which are capable of processing and analyzing complex textual and contextual data, can be used to improve the safety-critical assessment of driving scenarios.

Objective

The primary objective of this project is to develop a framework that uses LLMs for risk-aware analysis of generated driving scenarios. The framework will evaluate whether a scenario is safety-critical, provide feedback to refine and improve the scenarios, and test their effectiveness in validating motion planning and control algorithms. This approach aims to enhance the overall reliability of driving functions by ensuring a comprehensive exploration of accident-prone situations.

We Offer:

· A dynamic and future-oriented research environment
· Hands-on experience with a state-of-the-art software stack for autonomous driving
· Opportunity to publish a scientific paper (based on merit)
· The thesis can be written in either English or German

Requirements (What You Should Bring):

· Initiative and a creative, problem-solving mindset
· Excellent English or German proficiency
· Advanced knowledge of Python or C++
· Prior experience with autonomous vehicles or Large Language Models is an advantage
· Familiarity with common software development tools (e.g., Git, Ubuntu) is desirable

Work can begin immediately. If you are interested, simply send an e-mail with a brief cover letter explaining what fascinates you about this topic, a current academic transcript, and a resume to yuan_avs.gao@tum.de.

Kontakt: yuan_avs.gao@tum.de