Controllable Collision Scenario Generation via Collision Pattern Prediction

TL;DR

The paper tackles the problem of evaluating autonomous vehicle safety by generating controllable, safety-critical collision scenarios conditioned on a target collision type and time-to-accident ($TTA$). It introduces COLLIDE, a large-scale dataset balanced across five collision types and $TTA$ intervals, and a Collision Pattern representation that captures the ego–attacker configuration at impact. A region-proposal–inspired framework first predicts the Collision Pattern and then uses a quintic polynomial planner to realize feasible attacker trajectories, achieving higher collision rates and pattern accuracy than strong baselines. Experiments show that generated scenarios expose planner weaknesses across multiple planning systems and can be used to fine-tune planners for improved robustness, contributing to safer AV deployment in diverse collision scenarios.

Abstract

Evaluating the safety of autonomous vehicles (AVs) requires diverse, safety-critical scenarios, with collisions being especially important yet rare and unsafe to collect in the real world. Therefore, the community has been focusing on generating safety-critical scenarios in simulation. However, controlling attributes such as collision type and time-to-accident (TTA) remains challenging. We introduce a new task called controllable collision scenario generation, where the goal is to produce trajectories that realize a user-specified collision type and TTA, to investigate the feasibility of automatically generating desired collision scenarios. To support this task, we present COLLIDE, a large-scale collision scenario dataset constructed by transforming real-world driving logs into diverse collisions, balanced across five representative collision types and different TTA intervals. We propose a framework that predicts Collision Pattern, a compact and interpretable representation that captures the spatial configuration of the ego and the adversarial vehicles at impact, before rolling out full adversarial trajectories. Experiments show that our approach outperforms strong baselines in both collision rate and controllability. Furthermore, generated scenarios consistently induce higher planner failure rates, revealing limitations of existing planners. We demonstrate that these scenarios fine-tune planners for robustness improvements, contributing to safer AV deployment in different collision scenarios. Project page is available at https://submit-user.github.io/anon2025

Figures (3)

• Figure 1: Given an initial scenario, users specify condition inputs including collision type and time-to-accident (TTA). The framework predicts a Collision Pattern, defined as the relative configuration between the ego vehicle (green car) and the attacker (red bounding box) at the collision moment. This predicted pattern then guides the trajectory planner to generate a feasible attacker motion (red arrow), resulting in controllable generated scenarios such as lane change or LTAP.
• Figure 2: The proposed architecture for controllable collision scenario generation. Given structured scene context and user-specified conditions (collision type and TTA), our model predicts the “collision pattern”, representing the relative configuration of ego and attacker vehicles (ATT) at the moment of collision. The predicted pattern then guides a motion planner to produce a feasible attacker trajectory.
• Figure 3: Qualitative comparison of generated scenarios across different collision types. We compare our dataset and model against C-STRIVE rempe2022generating, the strongest baseline. Our models demonstrate strong controllability in generating user-specified collision types.