BUP48

The increasing complexity of automated vehicles (AVs) demands robust virtual validation environments that simulate diverse, real-world driving conditions. Existing scenario generation techniques often depend on structured templates or limited real-world data, which restricts the ability to capture rare or unexpected driving events crucial for AV validation. This project proposes a novel approach by leveraging Large Language Models (LLMs) to generate driving scenarios, improving the diversity and realism of AV validation environments.

Our approach is a two-step process. First, general LLMs will be utilized to design functional scenarios that describe a broad spectrum of driving conditions. In the second step, these abstract scenarios will be transformed into concrete executable scenarios, adhering to the standardized ASAM OpenSCENARIO and OpenDRIVE formats. This will facilitate their integration into virtual validation environments, allowing for seamless simulation execution.

By employing LLMs, we aim to automate and enhance scenario generation, addressing the limitations of current methods that lack scalability and diversity. This project not only promises to improve the accuracy and comprehensiveness of AV testing but also paves the way for more reliable and safer AV systems. Ultimately, this work supports the ongoing transformation towards automated, climate-friendly transportation systems.

Aim

The project aims to develop a novel framework using LLMs to generate realistic and diverse driving scenarios for testing automated vehicles. This approach seeks to enhance scenario variety beyond traditional methods, which are often limited by lack of diversity and reliance on predefined templates or incomplete real-world data.

Approach

The project employs a two-step methodology utilizing LLMs:

• Designing Abstract Scenarios: Using general LLMs series to generate a broad spectrum of potential driving conditions, including rare and complex situations.
• Concrete Scenario Development: Transforming these abstract scenarios into detailed, executable formats (e.g. ASAM OpenSCENARIO).

Benefit

This method provides significant innovations in automated vehicle testing:

• Enhanced Realism and Diversity: By generating scenarios that reflect a wider range of driving conditions, including unusual and unforeseen situations.
• Increased Efficiency: Streamlines the testing process by directly producing usable, detailed scenarios and related code, reducing the need for manual intervention.
• Scalability and Standard Compliance: Ensures that the scenarios and codes are not only realistic but also meet industry standards, making them immediately applicable for use in current simulation frameworks and adaptable for future technologies.