The DriveCoT dataset comprises 1058 scenarios and 36K labeled samples (similar to nuScenes), collected at a 2 Hz frequency, averaging 17 seconds per scenario. We partition the dataset into training, validation, and testing sets at a ratio of 70%, 15%, and 15%, respectively, yielding 25.3K training, 5.5K validation, and 5.5K testing samples. To prevent data leakage, we assign all data from the same scenario to the same set. Additionally, we ensure that the distribution of CoT aspects is similar across all splits.
The file strutures are shown below and you can click the (+) symbol to expand and see more details.

File Structure Download link

Citation

If you find the dataset helpful, please consider citing us.

    @article{Wang_2024_DriveCoT,
        title = {DriveCoT: Integrating Chain-of-Thought Reasoning with End-to-End Driving},
        author = {Wang, Tianqi and Xie, Enze and Chu, Ruihang and Li, Zhenguo and Ping, Luo},
        journal = {arXiv preprint arXiv:2403.16996},
        year = {2024}
    }