Reproduce “TrafficAlign: Aligning Large Language Models for Traffic Scenario Generation” on OpenResearch

About this paper

Zhi Tu, Liangkun Niu, Tianyi Zhang

Purdue University researchers developed TrafficAlign, an automated framework that aligns large language models with real-world traffic distributions to generate diverse, safety-critical scenarios for autonomous driving simulation. This approach synthesizes scenarios from driving videos, validates them using a domain-specific language, and fine-tunes LLMs, resulting in scenarios that induce significantly higher collision rates in AD models and improve AD model performance upon fine-tuning.

Key insights

• Aligning LLMs with real-world traffic distributions through supervised fine-tuning significantly enhances their ability to generate challenging, safety-critical scenarios that accurately reflect diverse geographic contexts.
• The use of a symbolic Domain-Specific Language (DSL) for validation effectively addresses LLM hallucination and semantic incompleteness, ensuring the quality and relevance of synthesized training data.
• Scenarios derived from empirically observed traffic patterns are more effective for both rigorous testing and performance improvement of autonomous driving models compared to unaligned or rule-based methods.

Summary by alphaXiv.