Reproduce “SceneDiffuser++: City-Scale Traffic Simulation via a Generative World Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shuhan Tan, John Lambert, Hong Jeon, Sakshum Kulshrestha, Yijing Bai, Jing Luo, Dragomir Anguelov, Mingxing Tan and 1 more
A generative world model named SceneDiffuser++ enables city-scale, trip-level traffic simulation by dynamically populating scenes with agents and realistically controlling traffic lights. This system generates diverse, long-duration synthetic driving miles, showing improved realism for agent insertion/removal and traffic light transitions compared to previous methods.
Key insights
- Employing "soft clipping" during inference is critical for stable training and generating realistic sparse tensors with diffusion models, preventing out-of-distribution values for non-existent entities.
- A trade-off exists between simulation reactivity (frequent replanning for better collision avoidance) and realism in dynamic agent generation (less frequent replanning for improved insertion/removal).
- Allowing dynamic agent insertion and removal enables the world model to generate more realistic agent behaviors and speeds, as it avoids the constraint of maintaining a fixed initial agent population.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.