TLDR: This paper introduces a GenAI-driven method for automating automotive software development, particularly for autonomous and ADAS features. It uses Large Language Models (LLMs) and Retrieval Augmented Generation (RAG) to transform requirements into test scenarios and code (Python for simulation, C++ for hardware). Model-Driven Engineering (MDE) is integrated for consistency checks, aiming to reduce development and testing time and improve compliance in the complex automotive industry.
The automotive industry, especially with the rise of Software Defined Vehicles (SDVs) and advanced driver-assistance systems (ADAS), faces immense challenges. Developing these complex systems involves managing hundreds of thousands of requirements, adhering to strict regulations, and enduring lengthy, costly development cycles. This often leads to a significant gap between initial research and full-scale production, heavily relying on specialized expertise and manual effort.
A new research paper, titled “GenAI for Automotive Software Development: From Requirements to Wheels,” introduces a groundbreaking approach to automate automotive software development using Generative AI (GenAI). Authored by Nenad Petrovic, Fengjunjie Pan, Vahid Zolfaghari, Krzysztof Lebioda, Andre Schamschurko, and Alois Knoll from the Technical University of Munich, this work focuses on streamlining the creation of autonomous and ADAS capabilities. You can find the full paper here: GenAI for Automotive Software Development.
The core idea is to transform initial requirements into functional software and test scenarios with minimal human intervention. The process begins with detailed requirements and generates crucial outputs like test scenario code for simulation environments and actual implementation code for vehicle hardware connected to a testbench. A key innovation is the integration of Model-Driven Engineering (MDE) for early consistency checks of requirements, helping to tackle issues like “hallucinations” from AI models—where the AI produces plausible but inaccurate information.
Large Language Models (LLMs) are central to this workflow. They are used for various tasks, including summarizing requirements into formal models, generating test scenarios, and producing both Python code for simulations and C++ code for the vehicle’s target platform. To further enhance the accuracy and relevance of generated test scenarios, Retrieval Augmented Generation (RAG) is employed. RAG helps by pulling information directly from autonomous driving regulations and related documents, ensuring compliance and precision.
The proposed methodology aims to significantly shorten compliance and re-engineering cycles, ultimately reducing the overall development and testing time for ADAS-related features. This is achieved through a comprehensive workflow that automates several critical steps.
Also Read:
- Decoding Driver Intent: AI Uncovers Reasons Behind Vehicle Braking
- A New Framework for Scaling Automated Driving Across Global Differences
How the GenAI Workflow Operates
The workflow starts by processing input documents, such as customer requirements or regulatory standards, using RAG. This step efficiently extracts relevant information, which then forms the basis for code and test generation. For instance, regulatory documents like UN Regulation No. 152 are used to generate test scenarios.
Next, LLMs are used to create a formal representation of the extracted requirements. This structured format allows for early design-time checks, including compliance verification. Once the requirements are validated for completeness and correctness, LLMs proceed to generate the necessary code. This includes Python code for simulating vehicle behavior in environments like CARLA and C++ code for the actual vehicle hardware on a testbench.
The paper details several key implementation components:
Model Checker: This component uses two LLM agents to generate model instances and constraints. It ensures consistency by checking Object Constraint Language (OCL) rules against the generated models. It leverages models like Llama 3.1-70B and a fine-tuned Llama3-8B.
Regulation-Compliant Scenario Generation: This crucial step focuses on creating precise, regulation-compliant test scenarios, especially for systems like Automated Emergency Braking Systems (AEBS). A robust two-stage RAG pipeline, featuring “SmartChunking” and “Smart Retrieve and Rerank,” is used to efficiently extract and process information from lengthy regulatory documents like UN Regulation No. 152.
Simulation Test Scenario Generation: Based on the regulation-compliant scenarios, this component generates configuration code for a CARLA-based simulation environment. Separate LLM-driven pipelines handle vehicle definition, pre-conditions (like scene setup and weather), and post-conditions (telemetry data and expected outcomes), primarily using GPT-4o.
Target Platform Code Generation: The final step involves generating C++ code for the target testbench platform, complementing the Python simulation code. This code handles vehicle control actions (steering, braking, acceleration) based on events from the simulation environment, integrating with vehicle signals and communication APIs.
In conclusion, this GenAI-powered approach for automotive engineering and re-engineering processes shows significant promise. By automating complex tasks and tackling the increasing complexity of SDV systems, it has the potential to drastically reduce the time needed for innovation and testing—from days or hours to mere minutes, particularly for automated emergency braking systems.
