CSG-Driver: Common Sense Guided Autonomous Driving under Legal Compliance and Practical Flexibility in Dilemma Situations

Yerin Lee, Seonggyu Lee, Shiho Kim

Keywords: Common Sense, Autonomous Driving, Large Language Models, Simulation-based Validation

TL;DR: This paper proposes CSG-Driver, a human-like autonomous driving system leveraging LLMs and common sense to handle long-tail scenarios by balancing traffic law compliance, safety, and adaptability.

Abstract:

Despite significant advancements in autonomous driving research, addressing long-tail cases remains a critical challenge. In this context, LLMs have gained attention for their interpretability and explainability, leading to increasing efforts to integrate them into autonomous driving tasks.

In this paper, we propose CSG-Driver, a human-like driving agent that combines compliance with road traffic laws and adaptability through the application of human common sense. We developed a closed-loop driving system within the CARLA simulator, which converts sensory data into natural language descriptions, incorporates road traffic laws, and utilizes prompts based on human driving behavior and past experiences. To address challenges in decision-making, the system employs common sense prompts and Chain-of-Thought reasoning to handle complex scenarios such as intersections with yellow lights, illegal parking avoidance, and highway driving.

Our experimental results demonstrate that CSG-Driver effectively resolves long-tail cases by leveraging LLMs to balance safety, traffic law compliance, and practical adaptability.

Submission Number: 22