Multi-Agent Large Language Model as AD/ADAS System Engineer

Abstract

Recent advancements in generative AI, particularly in Large Language Models (LLMs) have sparked a major revolution and a qualitative shift in various fields, including software code generation and unit test generation, offering new opportunities to automate various aspects of the software development process. At the same time, the demand of sophisticated software in the automotive industry has grown rapidly. This trend motivates the exploration of the potential of LLMs in supporting the development of AD/ADAS functions. A pipeline, CoTeGen, for code generation, test case generation, and the automation of virtual simulation-based testing in Esmini is designed following three iterative development cycles. The pipeline is designed to address four AD/ADAS functions. The first two are constrained to relatively elementary maneuvers, namely simple braking and lane changing, whereas the latter are dedicated to more sophisticated control tasks, specifically Adaptive Cruise Control and Collision Avoidance. Across these iterative cycles, the pipeline progressed from generating non-compilable software components to providing compilable and functional software. Based on a multi-run experimental evaluation involving five open-source LLMs, Codellama:7B, Mistral:7B, DeepSeek-Coder-v2:7B, Gemma3:4B, and Qwen2.5-Coder:7B, the pipeline shows a clear ability to generate correct source code for the simpler functions, while proving far less effective for the more advanced functions. Finally, we discuss the challenges and limitations of applying LLMs to code and unit test generation within the proposed pipeline.