Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
80 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
7 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Planning-oriented Autonomous Driving (2212.10156v2)

Published 20 Dec 2022 in cs.CV and cs.RO

Abstract: Modern autonomous driving system is characterized as modular tasks in sequential order, i.e., perception, prediction, and planning. In order to perform a wide diversity of tasks and achieve advanced-level intelligence, contemporary approaches either deploy standalone models for individual tasks, or design a multi-task paradigm with separate heads. However, they might suffer from accumulative errors or deficient task coordination. Instead, we argue that a favorable framework should be devised and optimized in pursuit of the ultimate goal, i.e., planning of the self-driving car. Oriented at this, we revisit the key components within perception and prediction, and prioritize the tasks such that all these tasks contribute to planning. We introduce Unified Autonomous Driving (UniAD), a comprehensive framework up-to-date that incorporates full-stack driving tasks in one network. It is exquisitely devised to leverage advantages of each module, and provide complementary feature abstractions for agent interaction from a global perspective. Tasks are communicated with unified query interfaces to facilitate each other toward planning. We instantiate UniAD on the challenging nuScenes benchmark. With extensive ablations, the effectiveness of using such a philosophy is proven by substantially outperforming previous state-of-the-arts in all aspects. Code and models are public.

Planning-oriented Autonomous Driving

The paper introduces Unified Autonomous Driving (UniAD), a framework designed to integrate perception, prediction, and planning tasks within autonomous driving. Unlike existing approaches that deploy standalone models or separate heads for these tasks, UniAD proposes a unified network that ties these components together with the aim of improving task coordination and minimizing accumulative errors.

The authors emphasize a planning-oriented design philosophy, suggesting that traditional modular designs cause significant information loss and error propagation. By utilizing a single network, UniAD seeks to ensure that all tasks are aligned towards the primary goal of effective planning.

Methodology

UniAD incorporates several modules, each performing specific tasks:

  • TrackFormer: This module handles detection and tracking, managing multi-object tracking (MOT) without the need for post-processing. It introduces detection and track queries to detect and track agents persistently, storing and updating the query state for continuous monitoring.
  • MapFormer: This performs online mapping via a panoptic segmentation approach to identify lanes, dividers, and crossings. Its results feed into the prediction modules to aid in understanding the driving environment.
  • MotionFormer: Responsible for forecasting agents' multimodal future movements, this module utilizes interaction modeling with agents and maps, and includes ego-vehicle queries for the self-driving car, factoring in spatial and temporal cues.
  • OccFormer: Conducts occupancy prediction by integrating dense scene features with agent-level knowledge, producing future occupancy maps that delineate agent identity and dynamics over time.
  • Planner: This component synthesizes the outputs from other modules to produce safe and reliable future trajectories, factoring in predicted occupancy to avoid potential collisions.

Experimental Insights

Extensive evaluation on the challenging nuScenes dataset demonstrates that UniAD not only outperforms previous multi-task solutions but also effectively combines previous standalone tracking, mapping, and predictive methods into a cohesive whole. Notably, UniAD achieves superior planning performance, reducing errors and demonstrating a notable decrease in planning-related collisions.

Implications and Future Work

The research outlines a significant step toward an integrated autonomous driving system that maximizes the utility of each component within the ecosystem. By unifying tasks into a cohesive network, UniAD reduces the complexity and interdependence issues present in previous modular approaches.

Future advancements could explore further optimization of UniAD's computational load, as well as extending the framework with additional tasks like depth estimation or advanced behavior prediction, potentially broadening its applicability across varied autonomous driving scenarios. Furthermore, customizing the system for different levels of autonomy and sensor configurations could enhance its adaptability in real-world applications.

This paper concludes by underscoring the potential of a planning-oriented approach in achieving a more reliable and interpretable autonomous driving solution, hinting at a transformative shift in how these systems can be designed and deployed moving forward.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (16)
  1. Yihan Hu (18 papers)
  2. Jiazhi Yang (8 papers)
  3. Li Chen (590 papers)
  4. Keyu Li (22 papers)
  5. Chonghao Sima (14 papers)
  6. Xizhou Zhu (73 papers)
  7. Siqi Chai (5 papers)
  8. Senyao Du (1 paper)
  9. Tianwei Lin (42 papers)
  10. Wenhai Wang (123 papers)
  11. Lewei Lu (55 papers)
  12. Xiaosong Jia (21 papers)
  13. Qiang Liu (405 papers)
  14. Jifeng Dai (131 papers)
  15. Yu Qiao (563 papers)
  16. Hongyang Li (99 papers)
Citations (425)
Github Logo Streamline Icon: https://streamlinehq.com

GitHub

  1. UniAD page (2,515 stars)