Scenic Test Scenario Generation and Repair for Autonomous Driving via Semantic Feedback and Knowledge Augmentation

Simulation-based scenario testing has become the core paradigm for autonomous driving system (ADS) safety validation, while real-world traffic accident reports, inherently safety-critical and scenically diverse, serve as a vital source for constructing high-quality test scenarios. However, existing methods for converting accident reports into executable Scenic test scenarios suffer from two critical deficiencies: the absence of a semantic feedback loop during code generation, and insufficient automated repair capability for low-resource domain-specific languages (DSLs). To address both issues, we propose SKGR (Semantic-feedback Knowledge-augmented Generation and Repair), a unified framework comprising a Semantic Feedback Correction (SFC) module that establishes a semantic-level verification loop for generated code, and a Dynamic Error Knowledge Base (DEKB) that constrains the repair search space via a three-tier knowledge architecture and dual-path retrieval, while a semi-automated asynchronous expansion mechanism mitigates cold-start barriers. Experiments conducted on 205 real-world California DMV accident reports demonstrate that the complete framework achieves a Compilation Pass Rate (CPR) of 92.5% and Semantic Accuracy (SA) of 85.9% on an independent test set, improving over the baseline by 4.2 and 8.6 percentage points respectively, with average repair iterations reduced by 40%.