Agentic generation of OpenX files from natural language descriptions

Usually, scenarios for testing of advanced driver assistance systems (ADAS) are generated utilizing certain scenario and road specification languages such as ASAM OpenSCENARIO and OpenDRIVE. Directly adopting these low-level languages limits the rate in which new scenarios are generated for virtual testing. Natural language (NL) would allow a much broader group of people and artificial intelligences to generate scenarios, increasing test coverage and safety. Instead of trying a direct translation from NL into OpenX, the existing intermediate domain specific language (DSL) stiEF is used. This not only facilitates testing and debugging but also generation, as its grammar can be used as a constraint for a large language model (LLM), which is then able to translate NL into stiEF. A parser is applied in an agentic way that interacts with the LLM until a syntactically correct file is generated, an optional second agent is then used to do basic semantic verification. Finally, the translation from a valid stiEF description into OpenX is done via conventional programming techniques.