Driving scenario generator
Driving scenario generator is an experimental tool, still in development, that uses a large language model to write driving scenarios, such as a car merging onto a highway, for testing automated driving systems. He worked on it in mid-2025, at the start of his research at JAIST on safety testing for automated driving.
How it works
A prompt is assembled from three kinds of reusable pieces, each kept as a file in the prompts folder:
- instruction tokens: what to generate;
- base scenarios: a scenario to start from (optional);
- experience tokens: lessons from earlier runs (optional).
A Python command-line tool combines them, sends the prompt to an OpenAI model (or to a mock client for testing without the API), and saves the result.
Status
The changelog is honest about the first version's limit: it produced descriptive scenarios, written as prose, which can't be fed directly to a simulator for testing. The structured, ontology-based side of the problem became his ADS safety-testing ontology and semantic extraction pipeline.