@article{Yu2026, 
author = {Haiyang Yu and Binghong Jiang and Xiaobo Qu and Li Li},
title = {Primitives and World Model for Road Transportation Foundation Models},
year = {2026},
journal = {Communications in Transportation Research},
keywords = {Transportation, foundation models, primitives, world model, large language models (LLM), tokens},
url = {https://www.sciopen.com/article/10.26599/COMMTR.2026.9640042},
doi = {10.26599/COMMTR.2026.9640042},
abstract = {Transportation foundation models are regarded as key infrastructure for achieving the leap from informatization to intellectualization in road transportation systems, and thus have become one of the most focused research directions among transportation researchers. Starting from first principles and based on the fundamental transportation data currently available, this paper abstracts two types of primitives as the core and fundamental data building blocks for transportation foundation models. Specifically, individual travel data are represented by vehicle trajectories as the first type of primitive data, with travel attribute data serving as labels for trajectory data. For data at specific spatial scales, traffic flow spatiotemporal series data of people or vehicles within a given time window are adopted as the second type of primitive data. Furthermore, this paper discusses how to construct a World Model for Transportation that generates primitive data representing future traffic states. This world model takes as input both sensor-acquired primitive data describing historical traffic states and time-series data of potential control and management actions (e.g., signal timing, route guidance, or lane closure plans), enabling it to simulate the outcomes of different intervention strategies. These primitive data will provide critical data foundations for reasoning, simulation, and decision-making in upper-level applications of transportation foundation models, serving as a bridge connecting data perception and intelligent decision-making.}
}