Fine-Tuning an LLM Agent for Automated OpenSCENARIO Generation and CARLA Execution

Summary

Beschreibung Motivation Scenario-based testing is essential for validating autonomous driving and ADAS functions. OpenSCENARIO enables standardized, reproducible, and shareable scenario definitions, and CARLA provides a scalable simulation environment for execution. However, authoring diverse OpenSCENARIO scenarios manually is time-consuming, error-prone, and hard to scale. Recent advances in LLM agents and domain fine-tuning make it feasible to generate structured artifacts (e.g., XML) automatically — but robustness is still challenging: outputs must be schema-valid, semantically consistent, and executable in CARLA. This thesis investigates a learning-based generation pipeline: fine-tune an LLM agent (and optionally combine it with retrieval and validation tools) to reliably generate novel, valid, and executable OpenSCENARIO scenarios at scale. Goal Develop an end-to-end toolchain to: • Build or curate a dataset of OpenSCENARIO scenarios (templates + variants + execution outcomes) • Fine-tune an LLM (or instruction-tune a smaller model) to generate valid .xosc scenarios from structured prompts • Integrate automatic XML/schema validation + rule checks + CARLA execution feedback • Export a scenario suite with metadata and quality metrics Expected Deliverables • LLM-based scenario generation agent (with inference script) • Fine-tuning dataset (or dataset generation pipeline) + training config • OpenSCENARIO scenario library (templates + generated scenarios) • CARLA execution runner + logging and evaluation reports Required Skills • Python, Git, basic software engineering • Interest in LLMs (fine-tuning, prompting, evaluation) • Familiarity with CARLA/OpenSCENARIO helpful but not mandatory Start Date Work can begin immediately. Please send me an email along with a current transcript of records and your resume, to: yuan_avs.gao@tum.de