Autonomous Driving Scenario Mining for Abnormal Trajectory Simulation

The patent creates synthetic copies of realistic driving scenarios by training a generative model on real driving data. Instead of collecting extensive real-world test miles, the system learns the distribution of normal driving behaviors from existing data and generates synthetic scenarios that replicate realistic driving conditions, including unusual but plausible situations. This copying approach maintains scenario realism while eliminating the need for extensive physical testing.

The patent performs preliminary analysis of real driving data to train the generative model before actual scenario generation. By pre-learning the characteristics of normal and unusual driving behaviors from historical data, the system prepares a robust generative model that can subsequently produce realistic scenarios without requiring additional real-world testing. This preliminary training phase captures essential driving pattern distributions for later synthetic generation.

The system synthesizes unusual driving scenarios by learning from the distribution of real driving data. The generative model captures rare but realistic driving patterns during training and can reproduce them synthetically, maintaining the ability to discover unusual scenarios without requiring extensive additional real-world testing. This copying mechanism preserves scenario diversity and realism while improving efficiency.

The patent modifies the approach from collecting more real-world data to transforming existing data parameters through generative modeling. By changing how data is utilized—from direct collection to synthetic generation based on learned distributions—the system maintains scenario quality while improving productivity. The generative model learns to generate scenarios with appropriate statistical properties without requiring proportional increases in real testing.

The patent creates a universal generative model that can produce diverse driving scenarios across the entire operational design domain from a single training dataset. Instead of requiring separate data collection systems for different scenario types, the generative model learns the joint distribution of driving conditions and can synthesize diverse scenarios including unusual situations. This multi-functional approach achieves scenario diversity without proportionally increasing data collection complexity.

The system copies the essential characteristics of diverse real-world driving scenarios through the generative model during training, then reproduces them synthetically. By learning the underlying distributions and relationships in the training data, the model can generate diverse scenarios without requiring continuous expansion of real-world data collection infrastructure. The copying process captures scenario diversity efficiently.