跳到主要导航 跳到搜索 跳到主要内容

Autonomous racing policy using safety-critical evolution in knowledge-fusion reinforcement learning

  • Xiaohui Hou
  • , Minggang Gan*
  • , Wei Wu
  • , Qianzhao Ma
  • , Jie Chen
  • *此作品的通讯作者
  • Beijing Institute of Technology
  • Beijing Institute of Technology
  • Minzu University of China
  • Chinese Academy of Engineering

科研成果: 期刊稿件 › 文章 › 同行评审

摘要

This study focuses on the development of an autonomous racing policy, encompassing knowledge fusion and trajectory planning for racing vehicles to expedite lap time reduction at the vehicle's handling limit. We propose a Safety-Critical Evolution in Knowledge-Fusion Reinforcement Learning (SCE-KFRL) controller, which integrates expert knowledge into the RL framework with two coupling approaches: 1) Tracking the expert driver's manipulation to derive a basic policy; 2) Learning an RL residual policy with experience-embedded reward shaping to enhance the basic policy. The SCE-KFRL controller intricately integrates a safety-critical Evolutionary Algorithm (EA) and an Adaptive Safety Modifier (ASM), which are tailored to harness critical operational data at the vehicle's handling limits and adjust unsafe actions without compromising the overall exploration. After training on track maps of varying difficulty, the proposed controller demonstrated superior performance compared to other baseline algorithms, thus verifying its effectiveness and extensibility. We believe this technology can be utilized in everyday driving scenarios, thereby enhancing efficiency in special conditions and stability in critical situations and expanding the application of autonomous driving.

源语言英语
期刊论文编号116528
期刊Applied Soft Computing
卷204
DOI
出版状态已出版 - 1月 2027
已对外发布是

学术指纹

探究 'Autonomous racing policy using safety-critical evolution in knowledge-fusion reinforcement learning' 的科研主题。它们共同构成独一无二的学术指纹。

引用此