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 synthetic corner case generation methods to address these challenges. Existing methods, particularly for planning tasks, are limited in scalability, customization, and contextual relevance.
We introduce SCP, a novel approach that leverages Large Language Models to generate scalable and customizable planning-specific corner cases. With sufficient world knowledge inside LLMs, SCP is data-agnostic and training-free, incorporating intention-driven context by explicit natural language commands.
We also present SCP-NuPlan, a benchmark that repurposes NuPlan's dataset into a diverse set of challenging corner cases, and SCP-LimSim, a flexible simulation tool for creating complex, customized corner cases via natural language descriptions. Our experiments show significant performance degradation in planning algorithms when tested on SCP-NuPlan, and the experiments of SCP-LimSim further demonstrate the customizability of diverse agent behaviors in response to detailed language inputs. We will release SCP-NuPlan and SCP-LimSim soon.
