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

Yanhao Wu; Haoyang Zhang; Fei He; Rui Wu; Yanhu Shan; Congpei Qiu; Liang Gao; Wei Ke

cs.ROJan 5, 2026

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

Authors: Yanhao Wu, Haoyang Zhang, Fei He, Rui Wu, Yanhu Shan, Congpei Qiu, Liang Gao, Wei Ke, +1 more

Organizations: School of Software Engineering, XJTU · Horizon Robotics · Shenzhen Loop Area Institute · University of Chinese Academy of Sciences

Abstract

Practical autonomous driving requires models that generalize by reasoning through spatial-temporal possibilities to exclude unsafe outcomes. While state-of-the-art (SOTA) methods use parallel planning architectures, they fail to explicitly couple speed decisions with agent behavior along the driving path, leading to suboptimal coordination. To address this, we propose a cascaded framework that transforms longitudinal planning from an independent prediction task into a path-conditioned reasoning process. On the model side, we introduce an anchor-based regression design that conditions longitudinal prediction on the lateral drive path, and reformulate longitudinal planning as 1D displacement prediction along the path. This reduces geometric uncertainty and sharpens the model's focus on interaction-driven dynamics. On the data side, we introduce a planning-oriented data augmentation strategy that simulates rare safety-critical events by programmatically inserting agents and relabeling longitudinal targets to enforce collision avoidance. Evaluated on the challenging Bench2Drive benchmark, our method achieves SOTA performance with a driving score of 89.07 and a success rate of 73.18%, demonstrating significantly improved coordination and safety. Further evaluation on Fail2Drive confirms strong generalization to rare edge cases where parallel formulations typically fail. Project page:https://yanhaowu.github.io/AlignDrive/.

Figures & tables

Figure 1: (a) Drive path (black), trajectory (blue), and longitudinal displacement (red). Path waypoints are sampled spatially, trajectory waypoints temporally, and displacements represent traveled distance along the path at fixed time intervals. (b) Comparison of E2E paradigms. Parallel planning predicts the drive path and longitudinal trajectory independently, which can lead to potential coordination inconsistencies. In the example on the right, the independently predicted longitudinal trajectory is collision-free, but applying its speed along a separately predicted lateral path could cause a collision. In contrast, our cascaded paradigm first predicts the drive path and then regresses path-conditioned longitudinal displacements. With the path prior, the model identifies the potential conflict and outputs shorter displacements, yielding to avoid collision. Perception inputs are omitted for clarity.

Figure 2: Overview of the proposed AlignDrive system, which consists of three components. The Drive Path Predictor refines queries through cross-attention with image features to encode the drive path, maps, and agents. The Planning-oriented Data Augmentation enriches scenarios by inserting additional agents and relabeling longitudinal displacements. Finally, the Longitudinal Planning Module predicts forward displacements along the drive path; combined with the path, these displacements yield the final trajectory that is both collision-aware and spatially consistent. On the right side of the figure, the black numbers denote the scores of predicted drive paths, while the red numbers represent the scores of the corresponding longitudinal planning for each drive path.

Figure 3: (a) Planning-oriented augmentation. Non-threatening agents are inserted at a distance with unchanged GT displacements, while threatening agents are placed nearby and cause adaptive shortening of GT displacements. (b) Representation encoding. Inserted agents are projected into future positions, transformed to corner representations, and encoded via a Fourier encoder (top). Reference points are sampled by displacement anchors and encoded with MLPs. For clarity, although multiple drive paths are predicted in practice, only one representative path is illustrated here (bottom).

Expert: PDM-Lite [ 1 ]
Method	Driving Score ( $\uparrow$ )	SR (%) ( $\uparrow$ )	Driving Efficiency ( $\uparrow$ )	Comfort ( $\uparrow$ )
SpaceDrive [ 17 ]	78.02	55.11	-	-
SimLingo [ 25 ]	86.02	67.27	259.23	33.67
Expert: Think2Drive [ 18 ]
VAD [ 16 ]	42.35	15.00	157.94	46.01
SparseDrive [ 28 ]	44.54	16.71	170.21	48.63

Table 1: Closed-loop results of planning in Bench2Drive. Bold and underlined numbers indicate the best performance within different expert groups.

Method		Ability (%) $\uparrow$
Method	Mean	Merging	Overtaking	Emergency Brake	Give Way	Traffic Sign
UniAD-Base [ 12 ]	15.55	14.10	17.78	21.67	10.00	14.21
VAD [ 16 ]	18.07	8.11	24.44	18.64	20.00	19.15
DriveTransformer-Large [ 15 ]	38.60	17.57	35.00	48.36	40.00	52.10
HiP-AD [ 29 ]	65.98	50.00	84.44	83.33	40.00	72.10
AlignDrive	70.06	75.00	75.56	75.00	50.00	74.74

Table 2: Multi-Ability Results in Bench2Drive.

Privileged Methods
Method	RGB	Lidar	Privilege	Language	In-Distribution			Generalization
					DS ( $\uparrow$ )	SR ( $\uparrow$ )	HM ( $\uparrow$ )	DS ( $\uparrow$ )	SR ( $\uparrow$ )	HM ( $\uparrow$ )
PlanT 2.0 [ 8 ]	$\times$	$\times$	✓	$\times$	87.8	85.0	86.4	73.3	58.0	64.8
PDMLite-F2D [ 1 ]	$\times$	$\times$	✓	$\times$	95.6	97.0	96.3	94.0	95.3	94.6
Multimodal Methods
Orion [ 6 ]	✓	$\times$	$\times$	✓	53.0	52.0	52.5	51.2	46.0	48.5

Table 3: Closed-loop results on Fail2Drive benchmark. The best results for RGB-only methods are bolded , and the best for Privileged/Multimodal methods are underlined .

Method	Parameters	Latency	Driving Score	Success Rate (%)
VAD-Base [ 16 ]	-	224.3 ms	42.35	15.00
DriveTransformer-Large [ 15 ]	646 M	221.7 ms	63.46	35.01
HiP-AD [ 29 ]	97.4 M	138.9 ms	86.77	69.09
AlignDrive	117.2 M	177.5 ms	89.07	73.18
AlignDrive-Small	83.7 M	124.5 ms	87.45	71.82

Table 4: Comparison of inference efficiency and driving performance. AlignDrive-Small is a lightweight variant with fewer decoder layers. Experiments are conducted on an RTX 3090 GPU.

Variant	LP	DP	DA	Driving Score $\uparrow$	Success Rate (%) $\uparrow$	Collision Rate (%) $\downarrow$
A				83.21	63.18	22.7
B	✓			84.85	65.45	19.5
C	✓	✓		85.82	66.81	16.3
D	✓		✓	86.54	68.92	15.7
E	✓	✓	✓	89.07	73.18	11.4

Table 5: Ablation study on AlignDrive components. LP: uses lateral path prediction to condition longitudinal planning; DP: formulates longitudinal planning as displacement regression along the drive path; DA: applies planning-oriented data augmentation

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Visualization of planning-oriented data augmentation. The top row shows non-threatening agents, while the bottom row shows threatening agents. Inserted synthetic agents are indicated with dashed boxes. Red points denote the ego vehicle’s original trajectory, and blue lines represent the adjusted longitudinal displacements after augmentation.

Hyperparameter	nuScenes	Bench2Drive
Displacement threshold $\delta$	0.5 m	0.5 m
Insertion probability $\alpha$	0.05	0.1
Safe distance $d_{safe}$	3.0 m	3.0 m
Ego trajectory horizon	3 s	3 s

Appendix

Table 6: Dataset-specific hyperparameters for nuScenes and Bench2Drive.

Method	Params	L2 ( m ) $\downarrow$				Collision (%) $\downarrow$
Method	Params	1 s	2 s	3 s	Avg.	1 s	2 s	3 s	Avg.
VAD-Base [ 16 ]	–	0.41	0.70	1.05	0.72	0.03	0.19	0.43	0.21
GenAD [ 37 ]	–	0.28	0.49	0.78	0.52	0.08	0.14	0.34	0.19
SparseDrive-S [ 28 ]	85 M	0.29	0.58	0.96	0.61	0.01	0.05	0.18	0.08
DriveTransformer-Large [ 15 ]	646 M	0.16	0.30	0.55	0.33	0.01	0.06	0.15	0.07
HiP-AD [ 29 ]	90 M	0.28	0.53	0.87	0.56	0.01	0.05	0.15	0.07

Appendix

Table 7: Open-loop planning evaluation results on the nuScenes validation dataset.

Method	LP	RE	DA	Driving Score $\uparrow$	Success Rate (%) $\uparrow$	Collision Rate (%) $\downarrow$
Decouple				83.21	63.18	22.7
LP + Original	✓			87.47	68.18	15.4
LP + Reencode	✓	✓		85.82	66.81	16.3
Full (AlignDrive)	✓	✓	✓	89.07	73.18	11.4

Appendix

Table 8: Ablation on longitudinal planning (LP), agent query decoding–re-encoding (RE), and planning-oriented data augmentation (DA). Decouple: no LP; LP + Original: LP with original agent queries; LP + Reencode: LP with decoded–re-encoded queries; Full (AlignDrive): LP + Reencode + DA.

Run	Driving Score $\uparrow$	Success Rate (%) $\uparrow$	Driving Efficiency $\uparrow$	Comfort $\uparrow$
Run 1	89.07	73.18	212.07	16.86
Run 2	87.80	71.36	207.85	15.25
Run 3	88.05	70.00	210.08	17.10
Average	88.30	71.50	210.00	16.40

Appendix

Table 9: Multiple simulation runs of AlignDrive on Bench2Drive benchmarks. Driving Score, Success Rate, Driving Efficiency, and Comfort are reported for each run along with the average.

Figure 5: Effect of planning-oriented data augmentation on planning performance. All augmented variants ( $p=0.1,0.2,0.3,0.4$ ) outperform the no-augmentation baseline.

Figure 6: Red points are predicted drive paths, while blue points show longitudinal planning outputs (trajectory waypoints for the baseline, displacement sequences for ours). Relevant vehicles are highlighted in green. The baseline collides with cross-traffic, while our method avoids it.

Figure 7: Comparison of Baseline (a) and Ours (b) in a pedestrian cut-in scenario. The baseline model fails to avoid the pedestrian, resulting in a collision, whereas our method promptly reacts and avoids the accident. The pedestrian is highlighted with a black dashed circle.

Explore similar work

MomADv2: Reliable Temporal Memory for End-to-End Autonomous Driving
Aug 24, 2026Ziying Song, Shengkai Zhang, Lin Liu +8Causality-Aware End-To-End Autonomous Driving Autonomous Driving
School of Artificial Intelligence (School of Software), Yanshan University · Beijing Jiaotong University · Nanyang Technological University +5
GraphWorld: Long-Horizon Planning with World Models for End-to-End Autonomous Driving
Jun 15, 2026Ziying Song, Caiyan Jia, Lin Liu +8Causality-Aware End-To-End Autonomous Driving Autonomous Driving
Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, School of Computer Science and Technology, Beijing Jiaotong University, Beijing, China
OWMDrive: Causality-Aware End-to-End Autonomous Driving via 4D Occupancy World Model
Jun 29, 2026Junjie Cheng, Ruiqi Song, Ye Wu +3Causality-Aware End-To-End Autonomous Driving Drives
The School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China · Waytous Inc., Qingdao 266109, China · The State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China +1

Aug 24, 2026, cs.CVJ/K move · Enter open · S save

MomADv2: Reliable Temporal Memory for End-to-End Autonomous Driving

Ziying Song, Shengkai Zhang, Lin Liu, Peiliang Wu+7

School of Artificial Intelligence (School of Software), Yanshan University · Beijing Jiaotong University · Nanyang Technological University+5

Causality-Aware End-To-End Autonomous Driving Autonomous Driving Long-Horizon Planning Long Refinement Trajectories

Open paper AskPDF arXiv

Long-horizon planning is critical for safe autonomous driving in complex scenarios. Existing methods improve planning continuity with temporal memory, but such memory may become invalid and mislead decisions when the driving command changes. Thus, selectively leveraging useful history while suppressing command-inconsistent memory remains a key challenge. To address this issue, we propose MomADv2, a reliable state-space memory framework for long-horizon end-to-end autonomous driving. At its core, MomADv2 introduces a Selective State-Space Planning Memory Query Module, which filters historical planning queries based on temporal continuity and command consistency, selects planning modes relevant to the current command, and models the evolution of planning intentions through a selective state-space mechanism. To further alleviate local trajectory deviations and error accumulation in long-horizon planning, we design a Flow-Matching Trajectory Residual Refiner. It learns a continuous residual correction field from the refined planning output to the expert trajectory, enabling fine-grained trajectory refinement while preserving the stability of anchor-based planning. Extensive experiments on closed-loop NAVSIM and Bench2Drive, as well as open-loop nuScenes, demonstrate that MomADv2 improves long-horizon planning consistency and reduces the average collision rate by 15.6% over MomAD under 6-second planning.

Abstract

Explore similar work

MomADv2: Reliable Temporal Memory for End-to-End Autonomous Driving

GraphWorld: Long-Horizon Planning with World Models for End-to-End Autonomous Driving

OWMDrive: Causality-Aware End-to-End Autonomous Driving via 4D Occupancy World Model