CARLA-GS: Unified Corner-Case Synthesis

The quest for robust autonomous driving safety hinges on effectively simulating rare, safety-critical interactions. Existing simulators often tackle corner-case generation in isolation, with diffusion models struggling to maintain spatiotemporal consistency and physical realism. This paper introduces CARLA-GS, a novel pipeline that addresses these limitations by unifying diverse generation components within a single, modular framework.

Bridging Semantic Reasoning and Physical Execution

CARLA-GS tackles the multi-faceted problem of corner-case synthesis by strategically decoupling, yet tightly coupling, its core modules. Starting with real driving data, it reconstructs an editable Gaussian scene, incorporating geometry-consistent constraints. This visual foundation is then leveraged by a multi-agent LLM, which performs scene-level reasoning to pinpoint risky interactions and generate high-level, intent-driven waypoint trajectories. Crucially, the low-level motion control is delegated to CARLA, ensuring kinematic and dynamic feasibility via a PID controller. This architecture allows for semantic understanding and physically executable motion to coexist, enhancing the realism and controllability of simulated scenarios.

Photorealistic Corner Cases with Spatiotemporal Fidelity

The framework's innovation lies in its ability to generate photorealistic, spatiotemporally consistent videos that align with both semantic intent and physically feasible motion. By re-projecting simulated vehicle states back into the Gaussian scene for ego-centric rendering, CARLA-GS achieves a high degree of visual fidelity. Experiments conducted on the Waymo Open Dataset demonstrate the system's capability for controllable corner-case generation, producing outputs that are both visually convincing and behaviorally sound. This advancement is critical for training and validating autonomous driving systems against the most challenging edge cases.