SurrealDriver: Designing LLM-powered Generative Driver Agent Framework based on Human Drivers' Driving-thinking Data

Ye Jin, Ruoxuan Yang, Zhijie Yi, Xiaoxi Shen, Huiling Peng, Xiaoan Liu, Jingli Qin, Jiayang Li, Jintao Xie, Peizhong Gao, Guyue Zhou, Jiangtao Gong

TL;DR

This paper introduces SurrealDriver, a novel framework that leverages human drivers' "driving-thinking" data—their verbalized thought processes during driving—to train LLM-powered autonomous driving agents. By using this natural language data as chain-of-thought prompts, SurrealDriver significantly improves driving safety and human-likeness compared to baseline LLM agents.

What Problem Does This Paper Address?

Large Language Models (LLMs) have shown impressive reasoning capabilities, making them attractive for embodied tasks like autonomous driving. However, LLMs lack inherent human embodiment and real-world experience, leading to suboptimal performance and sometimes implausible actions in complex driving scenarios. Traditional methods for training embodied agents, such as reinforcement learning with expert demonstrations, are often expensive and require translating human knowledge into code. This paper addresses the challenge of imbuing LLM-based driving agents with human-like decision-making and safety by directly incorporating natural language insights from human drivers.

Methods

The core of SurrealDriver is a framework designed around four key considerations:

  1. Basic Driving Pipeline: This involves perception (understanding the driving situation through decomposed parameters), decision-making (reasoning using LLMs), and control (executing atomic driving actions).
  2. Short-Term Driving Memory: This module stores recent driving behaviors to ensure continuity and enable the amalgamation of basic actions into complex maneuvers.
  3. Human-aligned Long-Term Driving Guidelines: This component uses the collected human "driving-thinking" data to guide the agent's planning, aligning its behavior with human drivers. A "CoachAgent" assesses the DriverAgent's actions and provides feedback based on these guidelines.
  4. Strict Safety Criteria: This is a crucial layer that enforces safety through redundancy mechanisms, preventing hazardous maneuvers.

The framework decomposes complex driving scenarios into atomic scene understanding units and atomic driving action units, making them digestible for LLMs. The "driving-thinking" data is used as chain-of-thought prompts to guide the LLM's reasoning process.

Datasets

The paper introduces a novel "driving-thinking" dataset collected from 24 human drivers (10 expert, 14 novice) during urban driving experiments. Drivers verbalized their thought processes while reviewing recorded driving footage. This dataset includes their natural language descriptions of decision-making, reasoning, and potential rationales behind their actions. This data, along with demographic and questionnaire information, forms the basis for training and evaluating SurrealDriver.

Key Insights and Ideas

  • Human "Driving-Thinking" as Chain-of-Thought: The paper demonstrates that verbalized human thought processes during driving can be effectively used as chain-of-thought prompts for LLMs, bridging the gap between LLM reasoning and embodied task execution.
  • Embodiment through Natural Language: Instead of relying solely on code-based demonstrations or reinforcement learning, SurrealDriver shows that natural language transcripts of human decision-making can directly inform and improve agent behavior.
  • Layered Safety and Guidance: The combination of strict safety criteria, short-term memory for continuity, and long-term human-aligned guidelines provides a robust approach to developing safer and more human-like driving agents.
  • Expertise Transfer via Language: Expert drivers' systematic patterns and reasoning can be distilled into language and then transferred to LLM agents, allowing them to learn complex behaviors without explicit coding.

Key Results

  • Safety Improvement: The framework with expert demonstration data significantly reduced collision rates by 81.04% compared to a baseline LLM agent.
  • Human Likeness: The framework increased human likeness by 50% as assessed by human evaluators.
  • Ablation Studies:
    • Removing safety criteria increased collision rates by 57.46%.
  • Removing short-term memory increased collision rates by 82.96%.
  • Removing long-term guidelines increased collision rates by 83.03%.
  • Human Evaluation: Participants rated the full SurrealDriver framework (with safety, memory, and guidelines) as significantly more human-like than a baseline without these components.

Limitations and Future Work

  • LLM Decision Latency: The paper notes that LLMs can have significant decision-making latency, requiring adjustments to simulation time steps to make them practical for real-time driving.
  • Data Collection Cost: While more cost-effective than traditional RL, collecting high-quality, natural language driving-thinking data still involves conducting driving experiments and transcribing interviews.
  • Generalization to More Complex Scenarios: While evaluated in urban driving, further work is needed to assess the framework's performance in more extreme or diverse driving conditions.
  • Real-World Deployment: The current evaluation is primarily in simulation. Future work would involve testing SurrealDriver in real-world driving scenarios.

Summary updated

Abstract

Leveraging advanced reasoning capabilities and extensive world knowledge of large language models (LLMs) to construct generative agents for solving complex real-world problems is a major trend. However, LLMs inherently lack embodiment as humans, resulting in suboptimal performance in many embodied decision-making tasks. In this paper, we introduce a framework for building human-like generative driving agents using post-driving self-report driving-thinking data from human drivers as both demonstration and feedback. To capture high-quality, natural language data from drivers, we conducted urban driving experiments, recording drivers’ verbalized thoughts under various conditions to serve as chain-of-thought prompts and demonstration examples for the LLM-Agent. The framework’s effectiveness was evaluated through simulations and human assessments. Results indicate that incorporating expert demonstration data significantly reduced collision rates by 81.04% and increased human likeness by 50% compared to a baseline LLM-based agent. Our study provides insights into using natural language-based human demonstration data for embodied tasks. The driving-thinking dataset is available at https://github.com/AIR-DISCOVER/Driving-Thinking-Dataset.
BibTeX
@inproceedings{jin2024surrealdriver,
  author = {Ye Jin and Ruoxuan Yang and Zhijie Yi and Xiaoxi Shen and Huiling Peng and Xiaoan Liu and Jingli Qin and Jiayang Li and Jintao Xie and Peizhong Gao and Guyue Zhou and Jiangtao Gong},
  title = {{SurrealDriver: Designing LLM-powered Generative Driver Agent Framework based on Human Drivers' Driving-thinking Data}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2024},
  doi = {10.1109/iros58592.2024.10802229},
}
Figure