Special Feature • Master's Thesis
Schema-Guided KG Construction and Retrieval for LLM-Based Automotive Test Scenario Generation
Officially registered • Implementation underway
Institution: Technical University of Munich (TUM)
Chair: Chair of Robotics, AI and Real-Time Systems (Prof. Alois Knoll)
Advisor: André Schamschurko
Overview
This thesis investigates schema-guided Knowledge Graph (KG) construction and retrieval strategies to improve LLM-based automated generation of test scenarios directly from complex automotive regulatory standards.
The domain focus is generating structured test scenarios (speed / load condition / post-condition tuples) derived from UN Regulation No. 152 (Advanced Emergency Braking Systems for M1 and N1 vehicles).
Research Objectives and Methodology
- Comparative Evaluation of Retrieval Strategies: Systematic comparison of flat vector retrieval, LightRAG, and HippoRAG 2 for generating test scenarios from UN Regulation No. 152, benchmarked against 59 annotated test scenarios using precision, recall, and F1-score.
- Schema-Guided Knowledge Graph Construction: Designing and automatically constructing a schema-guided knowledge graph that classifies regulatory parameters (scenario type, speed, load condition, subsequent condition) and bounds retrieval to a single scenario family, correcting retrieval failures seen in general-purpose graph construction.
- Automotive Compliance Grounding: Evaluating extraction accuracy across diverse scenario variations to ensure full compliance verification without hallucinations.
Living Project Updates
This page serves as the central living hub for updates regarding the thesis. Architecture diagrams, benchmark results, and publication materials will be added here as implementation progresses.