Safety-Critical Traffic Simulation with Guided Latent Diffusion Model (2505.00515v1)

Abstract: Safety-critical traffic simulation plays a crucial role in evaluating autonomous driving systems under rare and challenging scenarios. However, existing approaches often generate unrealistic scenarios due to insufficient consideration of physical plausibility and suffer from low generation efficiency. To address these limitations, we propose a guided latent diffusion model (LDM) capable of generating physically realistic and adversarial safety-critical traffic scenarios. Specifically, our model employs a graph-based variational autoencoder (VAE) to learn a compact latent space that captures complex multi-agent interactions while improving computational efficiency. Within this latent space, the diffusion model performs the denoising process to produce realistic trajectories. To enable controllable and adversarial scenario generation, we introduce novel guidance objectives that drive the diffusion process toward producing adversarial and behaviorally realistic driving behaviors. Furthermore, we develop a sample selection module based on physical feasibility checks to further enhance the physical plausibility of the generated scenarios. Extensive experiments on the nuScenes dataset demonstrate that our method achieves superior adversarial effectiveness and generation efficiency compared to existing baselines while maintaining a high level of realism. Our work provides an effective tool for realistic safety-critical scenario simulation, paving the way for more robust evaluation of autonomous driving systems.

• The paper proposes a Guided Latent Diffusion Model (LDM) leveraging a graph-based VAE and novel guidance objectives to efficiently generate realistic and adversarial safety-critical traffic scenarios.
• The framework improves upon existing methods by ensuring physical plausibility through a sample selection module and demonstrating superior adversarial effectiveness and generation efficiency.
• This approach offers significant potential to reshape autonomous vehicle testing methodologies by enabling cost-effective evaluation under diverse, challenging scenarios.

Safety-Critical Traffic Simulation with Guided Latent Diffusion Model

This paper presents a novel approach to safety-critical traffic simulation using the Guided Latent Diffusion Model (LDM), specifically designed to evaluate autonomous vehicle (AV) systems under rare and challenging scenarios. The significance of this research lies in its ability to generate physically realistic and adversarial traffic scenarios, addressing common shortcomings of existing methods, such as unrealistic scene generation, inefficiency, and low physical plausibility.

The authors propose an LDM framework that employs a graph-based variational autoencoder (VAE) for learning a compact latent space representation, effectively capturing multi-agent interactions while enhancing computational efficiency. Within this latent space, the diffusion model executes a denoising process to produce feasible traffic trajectories. To guide this process towards generating safety-critical scenarios, novel guidance objectives are introduced, which facilitate adversarial and controllable scenario simulation. Additionally, a sample selection module leveraging physical feasibility checks is developed to ensure the generated scenarios adhere to physical constraints.

The paper provides empirical evidence through extensive experiments on the nuScenes dataset, demonstrating the model's superiority in terms of adversarial effectiveness and generation efficiency compared to baseline methods. Notably, the proposed framework maintains a high degree of realism, essential for robust AV system evaluation.

Key Contributions

1. Efficient Adversarial Simulation: The LDM framework performs denoising within a compact latent space obtained from a graph-based VAE, resulting in improved computational efficiency. This feature contrasts sharply with traditional rule-based simulators, which demand considerable domain expertise and extensive manual parameter tuning.
2. Novel Guidance Mechanism: The inclusion of innovative guidance objectives within the diffusion model facilitates the generation of adversarial scenarios that are behaviorally realistic, thereby pushing AV systems toward identifying potential vulnerabilities and driving system improvements.
3. Enhanced Realism in Scenario Generation: By integrating a sample selection module based on physical feasibility checks, the framework ensures that generated scenarios are not only adversarially effective but also adhere to physical constraints, enhancing their applicability for large-scale AV evaluations.

Implications and Future Directions

The results illustrate that the proposed LDM framework can potentially reshape AV testing methodologies by providing a tool that mirrors real-world driving behaviors and adversarial interactions without the associated costs and time delays of physical testing. The model's ability to produce diverse and physically plausible scenarios underpins its effectiveness for safety-critical traffic simulation.

For future research, integrating LLMs and AI agents into the LDM framework could lead to more sophisticated controllability and performance in traffic simulations. Additionally, exploring cross-domain applications of the diffusion model might offer insights into improving simulation techniques across various autonomous systems and scenarios. Crucially, ensuring the model's adaptability to evolving traffic norms and behaviors will be pivotal in maintaining its relevance in dynamic urban environments.