DriveInsight

Jiankun Li, Hao Li, Jiangjiang Liu, Zhikang Zou, Xiaoqing Ye†, Fan Wang, Jizhou Huang†, Hua Wu, Haifeng Wang

Baidu Inc.

^† Corresponding author

arXiv link

Overview

We present a debugging and analyzing tool for black-box end-to-end autonomous driving models.

Overall architecture of our DriveInsight framework.

Evaluation

Closed-loop evaluation on the Town05 Long & Short benchmarks. Our method achieve a competitive driving score while also achieving the highest route completion. The sign * denotes that we exclude data from town05 in the training set.

Method	Modality	Reference	Training frames	Town05 Long DS	Town05 Long RC	Town05 Short DS	Town05 Short RC
LBC	C	CoRL 20	150K	12.3	31.9	31.0	55.0
Transfuser	C+L	TPAMI 22	150K	31.0	47.5	54.5	78.4
ST-P3	C	ECCV 22	150K	11.5	83.2	55.1	86.7
VAD	C	ICCV 23	3.0M	30.3	75.2	64.3	87.3
ThinkTwice	C+L	CVPR 23	2.2M	70.9	95.5	-	-
MILE	C	NeurIPS 22	2.9M	61.1	97.4	-	-
Interfuser	C	CoRL 22	3.0M	68.3	95.0	94.9	95.2
DriveAdapter	C+L	ICCV 23	2.0M	71.9	97.3	-	-
Ours	C+L	-	1.8M	66.6	100.0	95.3	99.2
Ours*	C+L	-	1.5M	64.4	100.0	93.2	95.8

Citation

If you find this project helpful, please consider citing the following paper:

@article{li2024exploring,
  title={Exploring the Causality of End-to-End Autonomous Driving},
  author={Jiankun, Li and Hao, Li and Jiangjiang, Liu and Zhikang, Zou and Xiaoqing, Ye and Fan, Wang and Jizhou, Huang and Hua, Wu and Haifeng, Wang},
  journal={arXiv preprint arXiv:2407.06546},
  year={2024}
}