# Goal-oriented Autonomous Driving

**Authors:** Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, Hongyang Li  
**arXiv:** [2212.10156](https://arxiv.org/abs/2212.10156)  
**Date:** 2022-12-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2212.10156v1/

## Key points

- The modern autonomous driving system is composed of modular tasks performed in a sequential order: perception, prediction, and planning.
- There are two main approaches currently used: deploying standalone models for individual tasks or designing a multi-task paradigm with separate heads.
- Accumulative error and negative transfer effects can be problematic with these approaches.
- The authors argue the need for an algorithm framework optimized for self-driving car planning.
- They introduce Unified Autonomous Driving (UniAD), a comprehensive framework that incorporates full-stack driving tasks in one network.
- UniAD leverages the advantages of each module and provides complementary feature abstractions for agent interaction from a global perspective.
- Tasks within UniAD are communicated using a unified query design to facilitate collaboration towards planning.
- UniAD outperforms previous state-of-the-art methods in all aspects according to extensive ablations on the nuScenes benchmark.
- The authors provide a full suite of codebase and models to support future research in the community.

## Summary

The modern autonomous driving system is composed of modular tasks that are performed in a sequential order, including perception, prediction, and planning. As technology advances and sensors and hardware improve, there is a growing demand for a system that can perform a wide range of tasks to achieve higher-level intelligence. Currently, there are two main approaches used: deploying standalone models for individual tasks or designing a multi-task paradigm with separate heads. However, these approaches may suffer from accumulative error or negative transfer effects. In this paper, the authors argue that it is essential to develop an algorithm framework that is optimized for the ultimate goal of self-driving car planning. To achieve this goal, they revisit the key components within perception and prediction and prioritize the tasks hierarchically so that all these tasks contribute to planning. In order to address the limitations of existing approaches, they introduce Unified Autonomous Driving (UniAD), which is a comprehensive framework that incorporates full-stack driving tasks in one network. UniAD is designed to leverage the advantages of each module and provide complementary feature abstractions for agent interaction from a global perspective. The tasks within UniAD are communicated using a unified query design to facilitate collaboration towards planning. The authors instantiate UniAD on the challenging nuScenes benchmark and conduct extensive ablations to evaluate its effectiveness. The results show that using this philosophy significantly outperforms previous state-of-the-art methods in all aspects. The authors also provide a full suite of codebase and models to facilitate future research in the community. Overall, this paper presents an innovative approach to autonomous driving by introducing UniAD as a comprehensive framework that integrates various driving tasks into one network. The experimental results demonstrate its superiority over existing methods, highlighting its potential impact on advancing autonomous driving technology.