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

**Authors:** Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Haoran Wu, Sifa Zheng  
**arXiv:** [2405.19620](https://arxiv.org/abs/2405.19620)  
**Date:** 2024-05-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2405.19620v2/

## Key points

- Traditional modular systems in autonomous driving handle perception, prediction, and planning tasks separately.
- End-to-end paradigms aim to unify these tasks into a single framework for optimization with a focus on planning.
- Current methods face challenges in performance and efficiency, especially in ensuring planning safety due to computational complexity and simplistic design.
- SparseDrive introduces a symmetric sparse perception module integrating detection, tracking, and online mapping tasks using a fully sparse representation of the driving scene.
- SparseDrive adopts a parallel design approach for motion prediction and planning tasks, treating planning as a multi-modal problem with a hierarchical planning selection strategy for safe trajectories.
- SparseDrive outperforms existing methods in task performance while achieving higher training and inference efficiency.

## Summary

In the field of autonomous driving, traditional modular systems have been widely used. These systems handle different tasks such as perception, prediction, and planning separately. However, this approach often leads to information loss and error accumulation across modules. In contrast, end-to-end paradigms aim to unify these tasks into a single framework that is fully differentiable. This allows for optimization with a focus on planning. Despite their potential benefits, current methods still face challenges in terms of performance and efficiency. This is especially true when it comes to ensuring planning safety due to the computational complexity of bird's eye view (BEV) features and the simplistic design of prediction and planning components. To address these limitations, a new paradigm called SparseDrive has been proposed. introduces a symmetric sparse perception module that integrates detection, tracking, and online mapping tasks using a model architecture that learns a fully sparse representation of the driving scene. Additionally, for motion prediction and planning tasks, a parallel design approach is adopted based on observations of similarities between these two tasks. The parallel motion planner in  models planning as a multi-modal problem and incorporates a hierarchical planning selection strategy with a collision-aware rescore module. This strategy aims to select rational and safe trajectories as final planning outputs. With its innovative designs,  significantly outperforms existing state-of-the-art methods in terms of task performance while also achieving higher training and inference efficiency. The authors Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Haoran Wu,and Sifa Zheng have provided code for  on GitHub to facilitate further research in the field of autonomous driving.