OmniSCS generates photorealistic safety-critical driving scenarios at 13Hz
New system edits agent states while preserving high-fidelity appearance and trajectory.
A team of researchers from multiple institutions has introduced OmniSCS (Omni Safety-Critical Scenario Synthesis), a novel system that generates photorealistic and physically accurate safety-critical scenarios for autonomous driving testing. The system comprises two key modules: a Fully Editable Driving World Construction module that maintains high-fidelity agent appearance and background during scene editing via dual-strategy agent reconstruction and depth-refinement background reconstruction methods, and an SCS Synthesis module that facilitates object insertion and agent trajectory editing to synthesize diverse scenarios while preserving data fidelity. Experimental results on the nuScenes, Waymo, and KITTI datasets demonstrate that OmniSCS significantly outperforms existing state-of-the-art methods in edited scene fidelity. The system supports real-time closed-loop testing at 13Hz, enabling efficient validation of autonomous driving algorithms. This work addresses a critical gap in current simulation tools that struggle to maintain data integrity after scene editing, offering a safer, more effective, and cost-efficient solution for safety-critical scenario optimization and testing.
- OmniSCS uses a Fully Editable Driving World module with dual-strategy agent reconstruction and depth-refinement background to maintain high fidelity after scene editing.
- The SCS Synthesis module enables object insertion and trajectory editing, allowing generation of diverse safety-critical scenarios while preserving data fidelity.
- Outperforms state-of-the-art methods on nuScenes, Waymo, and KITTI datasets, and supports real-time closed-loop testing at 13Hz.
Why It Matters
Enables safer, cheaper, and more scalable testing of autonomous driving systems by generating realistic edge cases.