Synthetic Driving Scene Generation Aligned With AV Test Goals

The patent replaces traditional manual or rule-based mechanical methods for generating synthetic driving scenarios with machine learning models (GANs and GPT). These models automatically generate diverse and realistic driving scenarios from text descriptions, eliminating the need for cumbersome manual configuration while simultaneously increasing scenario diversity and realism through learned patterns from training data.

The patent introduces text descriptions as an intermediary between the user and the complex scenario generation process. Users provide high-level text descriptions of desired scenarios, and the machine learning models automatically translate these into detailed scenario configurations with precise parameter control. This intermediary layer simplifies the user interface while maintaining precise control over scenario parameters through the models' learned understanding of driving contexts.

The patent employs pre-trained machine learning models that have already learned from extensive training data before deployment. When generating synthetic driving scenarios for validation, these pre-trained models can quickly produce diverse scenarios covering various operating goals without requiring time-consuming training at runtime. The preliminary training phase enables rapid generation during actual testing, achieving both comprehensive validation coverage and efficient generation times.