Services
IoT and Digital Engineering
Bringing Together IoT and Digital Engineering to Create Rich Experiences
Highlights
- Generative AI (GenAI) accelerates the transition to software-defined vehicles (SDVs), enabling rapid feature updates, personalized enhancements, and enhanced decision-making systems.
- By automating requirement analysis, improving design and development, and streamlining validation with AI-generated test cases and virtual environments, GenAI significantly reduces time-to-market while maintaining high accuracy and quality standards.
- The adoption of GenAI demands a fusion of advanced techniques, fine-tuned models, robust data curation, and scalable infrastructure to maximize performance, scalability, and ROI in the automotive domain.
On this page
Introduction
GenAI drives rapid innovation, enabling advanced features in SDVs.
The automotive industry is experiencing a major shift with the emergence of SDVs, where software plays a vital role in vehicle functionality and user experience. The ability to quickly iterate, update, and enhance software is essential in this new era. At the same time, the rise of GenAI is transforming software development processes. GenAI facilitates domain-specific feature development, including advanced decision-making systems for autonomous driving 2.0, conversational assistants, predictive diagnostics, and cybersecurity enhancements. This article explores how GenAI accelerates software development and supports the realization of software-defined vehicles.
GenAI for automotive
GenAI accelerates automotive software development, from requirement analysis to validation, enabling faster updates, enhanced quality, and efficient handling of complex systems.
The industry is shifting from traditional 3 to 5-year vehicle release cycles to an era of on-the-go feature updates. Vehicles are becoming "mobiles on wheels," with software features that can be dynamically added or removed. Given the complexity and personalization demands, vehicle software lines of code (LOC) are already around 100 million and are expected to grow. This shift requires faster software development and feature upgrades, often within weeks. Current processes and tools struggle to meet these timelines for large-scale software.
GenAI is propelling us from software 1.0 to software 2.0. In software 1.0, engineers focused on developing software from requirement specifications, validated against manually created test cases. In software 2.0, engineers focus on AI models that develop the software. GenAI enables teams to have AI companions that assist, augment, or transform each phase of the software development lifecycle, accelerating the entire process. This approach has transformative potential for the end-to-end automotive software development lifecycle, enhancing efficiency and software quality.
- GenAI for requirement analysis
Requirement engineering is a major challenge in the automotive industry, often hampered by skill shortages. GenAI can help create structured specification documents, decompose system requirements into subsystem and software-level requirements, and assist in various analysis methodologies like automatic decision table creation. Additionally, GenAI can aid in ISO 26262 functional safety-specific analyses, such as creating failure mode and effects analysis (FMEA) safety analyses.
- GenAI for design and development
GenAI enables rapid vehicle design and styling by transforming rough sketches into detailed models and allowing real-time modifications based on the sketch. While technology is not yet mature enough to generate production-ready code, GenAI can refactor code and improve code quality. For example, it can automatically fix static code analysis errors and ensure code adheres to automotive standards. In autonomous driving, foundational models enhance object detection and decision-making, with studies showing an 18% improvement in labeling accuracy using GenAI compared to regular convolutional neural network (CNN) models.
- GenAI for validation
Software validation consumes over one-third of development efforts in the automotive industry. It involves multiple testing levels, including unit, feature, system, bench, and vehicle testing. One of the most labor-intensive tasks is generating test cases from structured/unstructured system requirements. Combining multiple GenAI techniques with regular automation can help engineers generate test cases and scripts tailored to different domains and validation levels. This automation significantly reduces time and human effort while maintaining high accuracy and coverage.
Generative adversarial networks (GANs) can generate and synthesize large datasets from limited real-time data, creating virtual environments and simulating real-world scenarios quickly and cost-effectively. This allows autonomous vehicles to learn and adapt in controlled environments.
Proven strategies
Effective generative AI in automotive relies on proper methodologies, fine-tuning, data management, and strong infrastructure.
- Fusion-based approach: The success of GenAI in automotive applications depends on the choice of approaches to create a holistic solution. Various methodologies, including multiple types of prompt engineering, retrieval-augmented generation (RAG), and fine-tuning techniques, are available. There are over 20 prompt engineering techniques, 10+ RAG techniques like Native RAG and Graph RAG, and multiple fine-tuning techniques. Selecting the appropriate methodologies is essential for accelerating success, as they significantly impact the performance scalability and efficiency of AI solutions in the automotive sector.
Below is an approach we used to generate unit-level test cases based on requirements. We fused multiple prompt engineering techniques, RAG, LoRA-based LLM fine-tuning using the NVIDIA NeMoTM and NVIDIA NIMTM and regular automation. With this approach, we achieved approximately 90% accuracy and various structural coverage areas, such as 85% in decision, 87% in condition, and 73% in modified condition decision coverage (MCDC) for our target test dataset in the automotive body domain.
AI companion workflow
AI companion workflow
- Fine-tuning matters: Considering the long-term roadmap for this technology, models need to be configurable, fine-tunable, and scalable. Fine-tuning with automotive industry-specific data is critical for better results, allowing general-purpose models to specialize in particular tasks. This improves accuracy and ensures comprehensive coverage.
- Data curation is key: Proper data curation is vital to avoid degraded output due to improper weighting. Without strategic planning, overfitting or underfitting can occur, increasing the overhead of removing inaccurate results, review, and modification, thereby extending timelines instead of reducing effort. In the test case generation example, understanding the output patterns the model can and cannot generate was crucial for effective data curation for model fine-tuning.
- Build a robust GenAI ecosystem: A robust infrastructure and GenAI-based software platforms are essential. Partnerships with semiconductor and cloud providers are crucial. High-performance computing infrastructure with an appropriate infrastructure software stack is essential. For example, NVIDIA DGX cloud-based GPUs with NVIDIA NIMTM and NVIDIA NeMoTM software stacks, Amazon Bedrock, and Azure OpenAI can help manage large-scale data efficiently with excellent response times.
Conclusion
GenAI delivers value when applied strategically with the right expertise and infrastructure.
GenAI is not a one-size-fits-all solution. It is crucial to identify the problem before assessing whether GenAI can provide an effective solution. Given that the technology is still maturing, it is advisable to use existing tools that can deliver 100% accurate results when available.
The strategy and approach to developing GenAI solutions are crucial. Achieving successful outcomes requires AI expertise, domain knowledge, robust infrastructure, and, most importantly, the ability to integrate AI within the domain seamlessly. An AI companion can enhance your return on investment (ROI) by optimizing operational processes and expanding effectiveness. To experience the benefits and utilization of full capability, start immediately.
