SparseDrive: End-to-End Autonomous Driving via Sparse Scene Representation

April 1, 2026 · View on GitHub

https://github.com/swc-17/SparseDrive/assets/64842878/867276dc-7c19-4e01-9a8e-81c4ed844745

News

1 April, 2026: Our following work SparseDriveV2 is released. arxiv, Code.
17 March, 2025: SparseDrive is accepted by ICRA 2025.
24 June, 2024: We reorganize code for better readability. Code & Models are released.
31 May, 2024: We release the SparseDrive paper on arXiv. Code & Models will be released in June, 2024. Please stay tuned!

Introduction

SparseDrive is a Sparse-Centric paradigm for end-to-end autonomous driving.

We explore the sparse scene representation for end-to-end autonomous driving and propose a Sparse-Centric paradigm named SparseDrive, which unifies multiple tasks with sparse instance representation.
We revise the great similarity shared between motion prediction and planning, correspondingly leading to a parallel design for motion planner. We further propose a hierarchical planning selection strategy incorporating a collision-aware rescore module to boost the planning performance.
On the challenging nuScenes benchmark, SparseDrive surpasses previous SOTA methods in terms of all metrics, especially the safety-critical metric collision rate, while keeping much higher training and inference efficiency.

Overview of SparseDrive. SparseDrive first encodes multi-view images into feature maps, then learns sparse scene representation through symmetric sparse perception, and finally perform motion prediction and planning in a parallel manner. An instance memory queue is devised for temporal modeling.

Model architecture of symmetric sparse perception, which unifies detection, tracking and online mapping in a symmetric structure.

Model structure of parallel motion planner, which performs motion prediction and planning simultaneously and outputs safe planning trajectory.

Results in paper

Comprehensive results for all tasks on nuScenes.

Method	NDS	AMOTA	minADE (m)	L2 (m) Avg	Col. (%) Avg	Training Time (h)	FPS
UniAD	0.498	0.359	0.71	0.73	0.61	144	1.8
SparseDrive-S	0.525	0.386	0.62	0.61	0.08	20	9.0
SparseDrive-B	0.588	0.501	0.60	0.58	0.06	30	7.3

Open-loop planning results on nuScenes.

Method	L2 (m) 1s	L2 (m) 2s	L2 (m) 3s	L2 (m) Avg	Col. (%) 1s	Col. (%) 2s	Col. (%) 3s	Col. (%) Avg	FPS
UniAD	0.45	0.70	1.04	0.73	0.62	0.58	0.63	0.61	1.8
VAD	0.41	0.70	1.05	0.72	0.03	0.19	0.43	0.21	4.5
SparseDrive-S	0.29	0.58	0.96	0.61	0.01	0.05	0.18	0.08	9.0
SparseDrive-B	0.29	0.55	0.91	0.58	0.01	0.02	0.13	0.06	7.3

Results of released checkpoint

We found that some collision cases were not taken into consideration in our previous code, so we re-implement the evaluation metric for collision rate in released code and provide updated results.

Main results

Model	config	ckpt	log	det: NDS	mapping: mAP	track: AMOTA	track: AMOTP	motion: EPA_car	motion: minADE_car	motion: minFDE_car	motion: MissRate_car	planning: CR	planning: L2
Stage1	cfg	ckpt	log	0.5260	0.5689	0.385	1.260
Stage2	cfg	ckpt	log	0.5257	0.5656	0.372	1.248	0.492	0.61	0.95	0.133	0.097%	0.61

Detailed results for planning

Method	L2 (m) 1s	L2 (m) 2s	L2 (m) 3s	L2 (m) Avg	Col. (%) 1s	Col. (%) 2s	Col. (%) 3s	Col. (%) Avg
UniAD	0.45	0.70	1.04	0.73	0.66	0.66	0.72	0.68
UniAD-wo-post-optim	0.32	0.58	0.94	0.61	0.17	0.27	0.42	0.29
VAD	0.41	0.70	1.05	0.72	0.03	0.21	0.49	0.24
SparseDrive-S	0.30	0.58	0.95	0.61	0.01	0.05	0.23	0.10

Quick Start

Citation

If you find SparseDrive useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.

@inproceedings{sun2025sparsedrive,
  title={Sparsedrive: End-to-end autonomous driving via sparse scene representation},
  author={Sun, Wenchao and Lin, Xuewu and Shi, Yining and Zhang, Chuang and Wu, Haoran and Zheng, Sifa},
  booktitle={2025 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={8795--8801},
  year={2025},
  organization={IEEE}
}