CARLA-GS Decouples Representation, Reasoning, and Physics for AD Corner Cases

arXiv·medium signal

CARLA-GS targets autonomous-driving safety evaluation, which is dominated by rare, safety-critical interactions, by deliberately synthesizing corner cases. Its contribution is architectural: decoupling scene representation, reasoning, and physics simulation so each can be controlled independently, a design pattern relevant to anyone building agentic simulation or synthetic-scenario generation systems.