CaDRE: Controllable and Diverse Generation of Safety-Critical Driving Scenarios using Real-World Trajectories

TL;DR

This paper tackles the problem of generating safety-critical driving scenarios for autonomous vehicles by casting the task as a Quality-Diversity optimization problem. It introduces CaDRE, a framework that perturbs real-world trajectories within bounded limits, preserving realism while jointly optimizing for high safety-criticality and diverse scenario behavior through a grid-archived, black-box optimization process. The authors further enhance exploration with an Occupancy-Aware Restart mechanism and demonstrate superior sample efficiency and diversity across three real-world scenarios from nuPlan compared to RL- and sampling-based baselines. The work enables controllable retrieval of scenarios by desired measure values, offering a practical tool for robust AV evaluation and training, with potential extensions to multi-vehicle perturbations and road-context integration.

Abstract

Simulation is an indispensable tool in the development and testing of autonomous vehicles (AVs), offering an efficient and safe alternative to road testing. An outstanding challenge with simulation-based testing is the generation of safety-critical scenarios, which are essential to ensure that AVs can handle rare but potentially fatal situations. This paper addresses this challenge by introducing a novel framework, CaDRE, to generate realistic, diverse, and controllable safety-critical scenarios. Our approach optimizes for both the quality and diversity of scenarios by employing a unique formulation and algorithm that integrates real-world scenarios, domain knowledge, and black-box optimization. We validate the effectiveness of our framework through extensive testing in three representative types of traffic scenarios. The results demonstrate superior performance in generating diverse and high-quality scenarios with greater sample efficiency than existing reinforcement learning (RL) and sampling-based methods.

Figures (5)

• Figure 1: (a) Overview of the CaDRE framework. CaDRE utilizes a black-box optimization algorithm to explicitly optimize for high-quality and diverse safety-critical scenarios. (b) Illustration of the Occupancy-Aware Restart (OAR) mechanism, a general extension to QD algorithms to improve exploration efficiency.
• Figure 2: Coverage and QD score v.s. number of samples. Solid lines represent the mean, and the shaded area presents the standard deviation over 5 perturbed vehicles.
• Figure 3: Histograms of measure values. We visualize the final archive of the perturbed background vehicle with the highest QD score in the unprotected cross-turn. The solid line is the kernel density estimate of the true distribution. Note that the Gaussian kernel may introduce some distortions since the true distribution is bounded.
• Figure 4: Visualization of generated trajectories. The leftmost column shows the original unperturbed scenarios. The numbers below are the measure values $[m_1, m_2, m_3]$, representing the mean steering perturbation, impact time, and impact angle, respectively.
• Figure 5: Visualization of final archives. A darker color means a higher objective value. Transparent cells mean we cannot find scenarios. We visualize the perturbed vehicles that have the highest QD score for each scenario respectively.