Autonomous driving systems must reliably handle rare and complex scenarios, known as corner cases, to ensure safety in real-world conditions. However, directly collecting corner case data is often costly, time-consuming, and insufficiently scalable. As a result, researchers have turned to corner case generation methods to address these challenges. Existing methods for planning tasks face three fundamental limitations: (1) weak contextual relevance caused by environment-agnostic generation, (2) poor scalability due to manual scenario engineering, and (3) limited customizability from fixed behavioral templates.
We present SCP, a dual-paradigm framework that synergizes global scenario analysis and individual agent control through Large Language Models (LLMs). Our approach enables scalable generation via top-down coordination of agent trajectories, and customizable behaviors through bottom-up natural language directives, with contextual relevance by language-mediated spatial-temporal reasoning. SCP operates in a data-agnostic, training-free manner through two concrete instantiations.
One is SCP-NuPlan benchmark, with 6,382 challenging scenarios, causing 21.1% performance degradation in state-of-the-art planners. The other is SCP-LimSim simulator, enabling real-time behavior customization via natural language interaction. The proposed framework establishes a new paradigm for safety-critical scenario generation, effectively bridging the gap between algorithmic testing requirements and real-world driving challenges. We will release SCP-NuPlan and SCP-LimSim soon.