Reasoning to Regulate: Chain-of-Thought for Traffic Rule Understanding

2026-07-27

AI Summary

This work addresses the challenge that existing autonomous driving systems struggle to accurately interpret traffic rules that depend on contextual and spatial layouts, often oversimplifying them as end-to-end prediction tasks while neglecting the logical reasoning required for rule applicability. To bridge this gap, the study formulates traffic rule understanding as a vision-language reasoning problem and introduces a novel framework integrating a Chain-of-Thought (CoT) mechanism to explicitly link traffic sign semantics with high-definition map structures. The authors develop a scalable CoT annotation pipeline and propose an answer-oriented, fine-grained reward strategy, optimizing the system through large language model guidance, visual verification, supervised fine-tuning, and GRPO reinforcement learning. Evaluated on the MapDR dataset, the approach significantly improves rule comprehension accuracy and produces interpretable reasoning traces, establishing a new paradigm for regulation-aware autonomous driving.

Abstract

Understanding and complying with traffic regulations is a safety-critical requirement for autonomous driving, yet remains challenging due to the diversity and context dependence of traffic signage. Importantly, regulation understanding is not a simple recognition task, but a reasoning problem: whether a rule applies depends on interpreting the sign in relation to the spatial layout of lanes and scene context. To support such reasoning, MapDR provide fine-grained annotations that link each traffic sign's regulatory rules to the specific lanes they govern. Existing methods, however, largely treat this as direct sequence prediction, ignoring the underlying reasoning that connects sign semantics and map structure. To address this limitation, we explicitly incorporate reasoning into this task and propose a framework that equips vision-language models (VLMs) with chain-of-thought (CoT) capabilities. We first design a scalable CoT curation pipeline that bootstraps rationales from a strong LLM through a two-round strategy and employs a VLM-based verifier to filter out incorrect cases, yielding a high-quality set of (CoT, answer) pairs. Building on this foundation, we adopt a two-stage training scheme: supervised fine-tuning (SFT) to teach rationale-to-answer generation, followed by GRPO reinforcement learning with answer-grounded, fine-grained rewards to further improve final answer accuracy. Extensive experiments on MapDR show that our approach significantly improves both interpretability and accuracy, establishing the first reasoning-based framework for regulation-aware autonomous driving.

Problem

• traffic rule understanding
• autonomous driving
• context-dependent reasoning
• traffic signage
• regulation compliance

Innovation

• Chain-of-Thought
• Traffic Rule Understanding
• Vision-Language Models
• Reinforcement Learning
• Autonomous Driving