Controllable Collision Scenario Generation via Collision Pattern Prediction

---

This repository contains the official code for Controllable Collision Scenario Generation via Collision Pattern Prediction, a method for controllable collision scenario generation in autonomous driving.

Authors: Pin-Lun Chen, Chi-Hsi Kung, Che-Han Chang, Wei-Chen Chiu, Yi-Ting Chen
Affiliation: National Yang Ming Chiao Tung University

---

We introduce Collision Pattern, a compact and interpretable representation of the relative configuration between ego-attacker at the collision moment. Given a safe scenario, the user specifies collision type and time-to-accident (TTA) to predict collision pattern. This pattern guides the quintic motion planner to generate a feasible attacker trajectory that realizes the specified collision.

System Requirements

• Linux ( Tested on Ubuntu 18.04 )

• Python3 ( Tested on Python 3.8 )

• PyTorch ( Tested on PyTorch 1.8.0 )

• CUDA ( Tested on CUDA 11.1 )

• GPU ( Tested on Nvidia RTX3090Ti )

• CPU ( Tested on Intel Core i7-12700, 12-Core 20-Thread )

• NuScenes-api

Usage

Preprocessing

To preprocess our COLLIDE data into vectorized representation:

bash scripts/preprocessing_data.bash

Training

To train the condition collision scenario generation model with COLLIDE:

bash scripts/train.bash

Inference

To generate prediction results:

python test.py

Video visualization

To generate video result based on .csv files created in the inference stage:

python plot_and_metric.py

Citation

@article{chen2025controllable,
  title={Controllable Collision Scenario Generation via Collision Pattern Prediction},
  author={Chen, Pin-Lun and Kung, Chi-Hsi and Chang, Che-Han and Chiu, Wei-Chen and Chen, Yi-Ting},
  journal={arXiv preprint arXiv:2510.12206},
  year={2025}
}