StyleDrive
收藏资源简介:
StyleDrive数据集是一个大规模的真实世界数据集,专为个性化端到端自动驾驶(E2EAD)设计,包含近30,000个驾驶场景,并标注了风格偏好。数据集由清华大学智能系统与计算实验室(AIR)创建,旨在解决当前自动驾驶系统中缺乏大规模、真实世界数据集,且缺乏对个性化驾驶风格的细粒度标注的问题。数据集涵盖了广泛的城市和乡村驾驶环境,通过静态环境特征提取和动态上下文线索推断,为端到端自动驾驶系统的个性化提供基础。
The StyleDrive dataset is a large-scale real-world dataset specifically designed for personalized end-to-end autonomous driving (E2EAD), comprising nearly 30,000 driving scenarios annotated with style preferences. Developed by the Intelligent Systems and Computing Laboratory (AIR) at Tsinghua University, it aims to address the two critical gaps in current autonomous driving systems: the scarcity of large-scale real-world datasets and the lack of fine-grained annotations for personalized driving styles. The dataset covers diverse urban and rural driving environments, and provides a solid foundation for the personalization of end-to-end autonomous driving systems through static environmental feature extraction and dynamic contextual cue inference.
StyleDrive 数据集概述
数据集简介
- 名称:StyleDrive
- 目标:建立首个大规模真实世界数据集,用于个性化端到端自动驾驶(E2EAD)的基准测试
- 特点:
- 包含丰富的多样化驾驶偏好标注
- 使用静态道路拓扑和微调的视觉语言模型(VLM)提取上下文特征
- 通过行为分析、基于VLM的建模和人类验证获取客观和主观偏好标签
数据集构造
- 方法:提出统一的框架用于建模和标注个性化驾驶偏好
- 流程:
- 场景构建:基于静态道路拓扑和VLM提取细粒度场景特征
- 偏好标注:结合行为分析和人类验证
数据集内容
- 基准测试结果:包含多种模型的性能比较(AD-MLP、TransFuser、WoTE、DiffusionDrive等)
- 评估指标:NC、DAC、TTC、Comf.、EP、SM-PDMS
- 最佳表现模型:
- NC:DiffusionDrive-Style (97.81)
- DAC:WoTE-Style (93.44)
- TTC:WoTE-Style (93.70)
- Comf.:AD-MLP-Style (99.90)
- EP:TransFuser-Style (84.95)
- SM-PDMS:DiffusionDrive-Style (84.10)
获取方式
- 数据集下载:https://huggingface.co/datasets/Ryhn98/StyleDrive-Dataset
- 模型权重下载:https://huggingface.co/datasets/Ryhn98/StyleDrive-Dataset
相关资源
- 论文:https://arxiv.org/abs/2506.23982
- 项目主页:https://styledrive.github.io/
引用信息
bibtex @article{hao2025styledrive, title={StyleDrive: Towards Driving-Style Aware Benchmarking of End-To-End Autonomous Driving}, author={Hao, Ruiyang and Jing, Bowen and Yu, Haibao and Nie, Zaiqing}, journal={arXiv preprint arXiv:2506.23982}, year={2025} }
联系方式
- 联系人:Ruiyang Hao
- 邮箱:haory369@gmail.com

- 1StyleDrive: Towards Driving-Style Aware Benchmarking of End-To-End Autonomous Driving清华大学智能系统与计算实验室(AIR, Tsinghua University) · 2025年



