AlignDrive: Aligned Lateral-Longitudinal Planning
for End-to-End Autonomous Driving

Yanhao Wu^{1, 2}, Haoyang Zhang², Fei He², Rui Wu², Congpei Qiu^{1, 3}, Liang Gao¹, Wei Ke¹, Tong Zhang⁴,

¹Xi'an Jiaotong University ²Horizon Robotics ³Shenzhen Loop Area Institute ⁴University of Chinese Academy of Sciences

We develop an end-to-end driving framework, AlignDrive, that tightly couples lateral and longitudinal planning and enables effective planning-oriented data augmentation. On this page, we show that AlignDrive:

Demonstrates comprehensive capability in handling diverse driving scenarios.
Exhibits superior longitudinal planning in challenging, safety-critical situations.
Demonstrates superior lateral planning and obstacle avoidance performance.
Outperforms the baseline that performs independent lateral and longitudinal planning.

For videos, the red curve indicates the predicted drive path, while the blue curve corresponds to the predicted longitudinal displacement sequence along this drive path.

A. Comprehensive capabilities of AlignDrive

B. Closed loop evaluation in different scenarios

Vehicle Turning Route

Pedestrian Crossing

Opposite Vehicle Running Red Light

Enter ActorFlow

Construction Obstacle

Parked Obstacle Two-Ways

C. Qualitative comparison with the baseline

We provide a qualitative comparison between AlignDrive and the baseline. The baseline follows previous SOTA: planning lateral and longitudinal in parallel, without path-conditioned longitudinal planning or data augmentation; details are provided in the paper.

Compared to this baseline on the Bench2Drive benchmark, AlignDrive improves the driving score from 83.21 to 89.07, raises the success rate from 63.18% to 73.18%, and substantially reduces collision rates.