Reproduce “SAFE-SIM: Safety-Critical Closed-Loop Traffic Simulation with Diffusion-Controllable Adversaries” 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
Wei-Jer Chang, Francesco Pittaluga, Masayoshi Tomizuka, Wei Zhan, Manmohan Chandraker
Safe-Sim is a closed-loop traffic simulation framework utilizing guided diffusion models to create realistic and controllable safety-critical scenarios for autonomous vehicle planning algorithm evaluation. It effectively generates a diverse range of adversarial events, demonstrating improved collision rates and realism compared to baseline methods, along with fine-grained control over adversarial agent behaviors.
Key insights
- Diffusion models can be guided effectively to generate safety-critical scenarios by integrating adversarial terms during the denoising process, while maintaining realism.
- A balance between adversarial objectives (e.g., collision encouragement) and regularization terms (e.g., route adherence) is crucial for creating challenging yet realistic interactions.
- Partial diffusion allows for precise control over the characteristics of collisions, like point and angle, leading to significantly higher diversity in generated safety-critical events.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.