Multi-Agent Collaborative Decision-Making Using Small Vision-Language Models for Autonomous Driving

Autonomous vehicles face challenges in complex environments due to the computational inefficiency of large language models (LLMs) and the lack of multi-agent collaboration in existing decision-making approaches. This paper proposes a small Vision-Language Model (VLM)-based framework for connected autonomous vehicles (CAVs), balancing computational efficiency with robust reasoning and perception. The framework integrates a perception module with mapping and annotation layers to provide visual and contextual information. To enhance decision-making, we introduce Domain-Augmented Chain-of-Thought (DACoT), enabling context-aware planning and negotiation for multi-agent coordination. In intersection scenarios, DACoT improves Success Steps (SS) by 28.9% and Pass Rate (PR) by 24.2% over CoT, and SS by 57.2% and PR by 48.2% over Direct prompting. In highway merging, DACoT achieves 19.6% higher SS and 99.97% higher PR than CoT. These results highlight DACoT’s adaptability in complex multi-agent interactions, establishing an efficient, scalable, and collaborative decision-making framework for real-time autonomous driving.