Robust Translation of Natural Language into Traffic Sequence Charts and Co-Simulation with Human Driver Models

Kurzfassung

Formal scenario descriptions are essential for verification, testing, and knowledge representation in automated driving, but domain experts often describe traffic situations in natural language. Translating such descriptions into formal scenario models is challenging because natural language is ambiguous, variable, and often underspecified, while direct end-to-end generation of formal artifacts can be brittle and difficult to validate. For this we present NL2TSC, a neural-symbolic pipeline for translating natural-language traffic scenario descriptions into formal Traffic Sequence Charts (TSCs). The approach uses reusable intermediate Spatial View Cards (SV-Cards) instead of unconstrained generation, enabling structured grounding in spatial configurations and TSC fragments. We position it as a robustness-oriented interface between informal scenario descriptions and formal traffic specifications. Further we want to show an extended pipeline to co-simulate concrete scenarios with human models. Using overtaking as an initial use case, we outline how the proposed design can support robust scenario formalization and its simulation.