Zhou, Shanghang (Tongji University), Yang, Shuo (Tongji University), ma, zhenyu (Tongji university), Du, Wei (Jilin University), Li, Xincheng (Tongji University), Xing, Jiaming (Tongji University), Huang, Yanjun (Tongji University)

A Comparative Study of Multi-Vehicle Knowledge Fusion Methods

Scheduled for presentation during the Regular Session "S10c-Cooperative and Connected Autonomous Systems" (WE-LA-T10), Wednesday, November 19, 2025, 16:40−17:00, Cooleangata 4

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 19, 2025

Abstract

Self-learning techniques show great potential to improve the adaptability of autonomous vehicles in dynamic environments. However, existing self-learning approaches mainly rely on isolated learning paradigms, resulting in fragmented knowledge accumulation. To expand the scope of self-learning techniques and strengthen the generalization capabilities of autonomous driving agents, many studies have explored multi-vehicle knowledge fusion, leveraging the knowledge accumulated across multiple agents to build more robust and scalable models. This paper reviews current knowledge fusion techniques and presents a case study of two core approaches: policy distillation and model merging. We evaluate these methods in representative traffic scenarios and discuss their respective strengths, limitations, and unique characteristics. Experimental results demonstrate that policy distillation delivers high accuracy at the cost of substantial resources, whereas model merging is more resource-efficient but less precise. By integrating both methods, we drastically reduce computational costs while achieving 92% of expert model performance in diverse environments, effectively combining the complementary strengths of data-driven and model-based approaches.