Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AdvDiffuser: Generating Adversarial Safety-Critical Driving Scenarios via Guided Diffusion

Yuting Xie, Xianda Guo, Cong Wang, Kunhua Liu, Long Chen

TL;DR

Safety evaluation of autonomous vehicles requires scalable long-tail driving scenarios that are rare in real data. AdvDiffuser introduces a guided diffusion framework that learns realistic multi-agent background traffic from offline logs and uses a reward-guided sampler to generate adversarial scenarios that challenge planners. The approach decouples realism from adversarial objectives, enabling transferability across unknown target planners with minimal online warm-up, and demonstrates superior realism, diversity, and adversarial effectiveness on nuScenes compared to existing baselines. This work highlights practical implications for robust AV testing and emphasizes the need for benchmarks that reflect the real-world likelihood of hazardous scenarios.

Abstract

Safety-critical scenarios are infrequent in natural driving environments but hold significant importance for the training and testing of autonomous driving systems. The prevailing approach involves generating safety-critical scenarios automatically in simulation by introducing adversarial adjustments to natural environments. These adjustments are often tailored to specific tested systems, thereby disregarding their transferability across different systems. In this paper, we propose AdvDiffuser, an adversarial framework for generating safety-critical driving scenarios through guided diffusion. By incorporating a diffusion model to capture plausible collective behaviors of background vehicles and a lightweight guide model to effectively handle adversarial scenarios, AdvDiffuser facilitates transferability. Experimental results on the nuScenes dataset demonstrate that AdvDiffuser, trained on offline driving logs, can be applied to various tested systems with minimal warm-up episode data and outperform other existing methods in terms of realism, diversity, and adversarial performance.

AdvDiffuser: Generating Adversarial Safety-Critical Driving Scenarios via Guided Diffusion

TL;DR

Safety evaluation of autonomous vehicles requires scalable long-tail driving scenarios that are rare in real data. AdvDiffuser introduces a guided diffusion framework that learns realistic multi-agent background traffic from offline logs and uses a reward-guided sampler to generate adversarial scenarios that challenge planners. The approach decouples realism from adversarial objectives, enabling transferability across unknown target planners with minimal online warm-up, and demonstrates superior realism, diversity, and adversarial effectiveness on nuScenes compared to existing baselines. This work highlights practical implications for robust AV testing and emphasizes the need for benchmarks that reflect the real-world likelihood of hazardous scenarios.

Abstract

Safety-critical scenarios are infrequent in natural driving environments but hold significant importance for the training and testing of autonomous driving systems. The prevailing approach involves generating safety-critical scenarios automatically in simulation by introducing adversarial adjustments to natural environments. These adjustments are often tailored to specific tested systems, thereby disregarding their transferability across different systems. In this paper, we propose AdvDiffuser, an adversarial framework for generating safety-critical driving scenarios through guided diffusion. By incorporating a diffusion model to capture plausible collective behaviors of background vehicles and a lightweight guide model to effectively handle adversarial scenarios, AdvDiffuser facilitates transferability. Experimental results on the nuScenes dataset demonstrate that AdvDiffuser, trained on offline driving logs, can be applied to various tested systems with minimal warm-up episode data and outperform other existing methods in terms of realism, diversity, and adversarial performance.

Paper Structure

This paper contains 15 sections, 12 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Testing Autonomous Driving in Simulation
  4. Safety-critical Driving Scenario Generation
  5. Diffusion Models for Sequence Decision Problems
  6. AdvDiffuser
  7. Problem Formulation
  8. Diffusion Model for Multi-agent Trajectories
  9. Reward Model for Guided Sampling
  10. Generating Safety-critical Scenarios
  11. Experiments
  12. Simulating Real Traffic
  13. Generating Accident-Prone Scenarios
  14. Transferability Analysis
  15. Discussion

Figures (3)

  • Figure 1: AdvDiffuser generates safety-critical scenarios for testing AV systems. (Top) Diffusion-based traffic simulation involves noise undergoing k rounds of reverse diffusion process. (Bottom) The diffusion model, coupled with a reward guide, generates adversarial background trajectories. AVs are depicted in green, while vehicles executing attack behaviors are in orange.
  • Figure 2: Overview of our AdvDiffuser framework. Given a context comprising historical vehicle trajectories and the map, AdvDiffuser aims to generate adversarial trajectories for background vehicles. The Interaction Net is a Graph Neural Network (GNN) architecture that encodes inter-vehicle interactions, providing a latent code $z_{prior}$ as a condition for the denoising process. The dynamics model $f$ in Trajectory Decoder transforms vehicle actions into trajectories.
  • Figure 3: Evaluation of few-shot learning. AdvDiffuser exhibits increased collision rates when encountering novel targeted planners following a few initial interactive steps.