Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Human-Centric Autonomous Driving: A Fast-Slow Architecture Integrating Large Language Model Guidance with Reinforcement Learning

Chengkai Xu, Jiaqi Liu, Yicheng Guo, Yuhang Zhang, Peng Hang, Jian Sun

TL;DR

The proposed architecture not only reduces collision rates but also aligns driving behaviors more closely with user preferences, thereby achieving a human-centric mode and bridges the gap between individual passenger needs and the rigor required for safe, reliable driving in complex traffic environments.

Abstract

Autonomous driving has made significant strides through data-driven techniques, achieving robust performance in standardized tasks. However, existing methods frequently overlook user-specific preferences, offering limited scope for interaction and adaptation with users. To address these challenges, we propose a "fast-slow" decision-making framework that integrates a Large Language Model (LLM) for high-level instruction parsing with a Reinforcement Learning (RL) agent for low-level real-time decision. In this dual system, the LLM operates as the "slow" module, translating user directives into structured guidance, while the RL agent functions as the "fast" module, making time-critical maneuvers under stringent latency constraints. By decoupling high-level decision making from rapid control, our framework enables personalized user-centric operation while maintaining robust safety margins. Experimental evaluations across various driving scenarios demonstrate the effectiveness of our method. Compared to baseline algorithms, the proposed architecture not only reduces collision rates but also aligns driving behaviors more closely with user preferences, thereby achieving a human-centric mode. By integrating user guidance at the decision level and refining it with real-time control, our framework bridges the gap between individual passenger needs and the rigor required for safe, reliable driving in complex traffic environments.

Towards Human-Centric Autonomous Driving: A Fast-Slow Architecture Integrating Large Language Model Guidance with Reinforcement Learning

TL;DR

The proposed architecture not only reduces collision rates but also aligns driving behaviors more closely with user preferences, thereby achieving a human-centric mode and bridges the gap between individual passenger needs and the rigor required for safe, reliable driving in complex traffic environments.

Abstract

Autonomous driving has made significant strides through data-driven techniques, achieving robust performance in standardized tasks. However, existing methods frequently overlook user-specific preferences, offering limited scope for interaction and adaptation with users. To address these challenges, we propose a "fast-slow" decision-making framework that integrates a Large Language Model (LLM) for high-level instruction parsing with a Reinforcement Learning (RL) agent for low-level real-time decision. In this dual system, the LLM operates as the "slow" module, translating user directives into structured guidance, while the RL agent functions as the "fast" module, making time-critical maneuvers under stringent latency constraints. By decoupling high-level decision making from rapid control, our framework enables personalized user-centric operation while maintaining robust safety margins. Experimental evaluations across various driving scenarios demonstrate the effectiveness of our method. Compared to baseline algorithms, the proposed architecture not only reduces collision rates but also aligns driving behaviors more closely with user preferences, thereby achieving a human-centric mode. By integrating user guidance at the decision level and refining it with real-time control, our framework bridges the gap between individual passenger needs and the rigor required for safe, reliable driving in complex traffic environments.
Paper Structure (29 sections, 17 equations, 7 figures, 1 table)

This paper contains 29 sections, 17 equations, 7 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. Reinforcement Learning for Autonomous Driving
  4. Human-Centric Autonomous System
  5. Hierarchical Decision Architectures
  6. Problem Formulation
  7. Action Space
  8. Observation Space
  9. Reward Function
  10. Methodology
  11. Framework overview
  12. LLM-Based Slow System
  13. Scene Encoding
  14. Memory Retrieval
  15. Instruction-Based Prompt Construction and Reasoning
  16. ...and 14 more sections

Figures (7)

  • Figure 1: The LLM interprets high-level human intent and generates structured guidance, while the RL module integrates this guidance with environmental context to make optimal, human-aligned decisions.
  • Figure 2: The proposed fast-slow architecture combines LLM guidance with RL. LLM interprets human intentions and converts them into structured instructions, while RL combines environmental information to make safe and robust decisions.
  • Figure 3: Overview of the embedding matrix of observation-instruction and the structure of the multi-head attention based policy network.
  • Figure 4: Three simulation scenarios designed to evaluate the framework’s capabilities from distinct perspectives: (a) high-speed driving, (b) right-lane vehicle merging, and (c) overtaking on a two-way road.
  • Figure 5: Comparison of the success rates of the proposed model and baselines model in three test scenarios.
  • ...and 2 more figures