Automatic Generation of vTESTstudio Test Cases from Natural Language Requirements Using Large Language Models

Abstract

Automotive testing still involves substantial manual effort when natural-language requirements are translated into tool-compatible test artefacts. As vehicle functions, carlines and software releases increase, this translation step becomes a growing bottleneck in test development. This thesis investigates a two-phase pipeline for generating vTESTstudio-compatible test artefacts from natural-language automotive requirements using Large Language Models (LLMs). The first phase converts a requirement into a structured intermediate representation of logical test steps. The second phase grounds these steps in valid domain resources, including signals and reusable functions, before constructing the final Vector Test Table (VTT) artefact. This separation makes the generation process easier to inspect, evaluate and control. The thesis further studies retrieval-based grounding, parameter-efficient fine-tuning for intermediate representation generation, and retrieved skill guidance for improving logical planning. On the small evaluation sets used in this thesis, the proposed pipeline produced useful test artefacts in selected cases, but human review remained necessary. This is assessed up to the generated logical test steps and the coverage of the VTT artefacts against reference cases, not their syntactic validity or execution in vTESTstudio or CANoe. Within these limits, retrieval tended to improve grounding indomainness and signal coverage, while fine-tuning improved the validity, consistency and domain style of intermediate representations. Retrieved skills helped planningoriented aspects such as logical adequacy and structural quality, though larger skill contexts could make downstream grounding harder. Overall, the thesis suggests that requirement-driven automotive test generation is more controllable when requirement interpretation, domain grounding and artefact construction are treated as separate stages.