Using local large language models to create more human-like agents in traffic simulation

Overview

Microscopic traffic simulation is widely used to study congestion, road safety, traffic flow, and interactions between human-driven and automated vehicles. However, many traffic simulators rely on rule-based driver behaviour models, such as car-following and lane-changing models, which may not fully capture the diversity, variability, and context-awareness of human driving.

This project will investigate whether local large language models can be used to create more human-like agents in traffic simulation. The project will build on an earlier student capstone study that explored the impact of replacing one vehicle in a simple traffic simulation with an LLM-driven vehicle. This VRES project will extend the idea toward a more scalable simulation environment using SUMO and Python.

Research engagement

This is exploratory research project at the intersection of traffic simulation and artificial intelligence. The project will focus on how large language models could influence the behaviour of selected vehicles in a traffic simulation environment.

The student will engage with:

• Literature on traffic simulation, driver behaviour modelling, agent-based simulation, and large language models.
• Microscopic traffic simulation using SUMO.
• Python-based control of simulated vehicles.
• Concepts of human-like driving behaviour, including cautious, aggressive, distracted, cooperative, or impatient driving styles.
• Critical reflection on the feasibility, limitations, and risks of using LLMs in traffic simulation.

Research activities

the main activities:

• Conduct a focused literature review on the use of LLMs, AI agents, and human-like behaviour models in traffic simulation.
• Review conventional driver behaviour models used in microscopic traffic simulation, including speed-based models, car-following models, and lane-changing models.
• Learn how to use SUMO and Python to run simple traffic simulations and control selected vehicles.
• Develop a simple architecture where a local LLM acts as a behavioural decision or a control layer for one or more selected vehicles.
• Explore how different driver profiles or behavioural instructions can affect simulated vehicle decisions.
• Compare selected LLM-augmented vehicle behaviours against conventional SUMO vehicle behaviours.
• Test the approach in simple traffic scenarios, such as following, merging, lane changing, or intersection interaction.
• Document the technical challenges, including latency, scalability, behavioural consistency, safety, and reproducibility.

Research skills

You will gain experience in:

• Conducting and synthesising a focused academic literature review.
• Understanding microscopic traffic simulation and driver behaviour models.
• Using or learning SUMO for traffic simulation.
• Using Python to interact with simulation environments.
• Exploring how local LLMs can be integrated into technical workflows.
• Designing simple agent-based simulation experiments.
• Comparing baseline and experimental simulation behaviours.
• Critically assessing the opportunities and limitations of emerging AI methods in transport research.
• Communicating research findings, technical assumptions, and future research directions.

Outcomes

The expected outcomes of the project are:

• A structured literature review on LLMs, AI agents, and human-like behaviour modelling in traffic simulation.
• A simple SUMO–Python prototype showing how a local LLM can influence the behaviour of selected simulated vehicles.
• A set of demonstration scenarios comparing conventional traffic simulation with LLM-augmented vehicle behaviour.
• An initial assessment of whether LLM-informed agents can produce more diverse, interpretable, or human-like simulated traffic behaviour.
• A short research report summarising the methodology, findings, limitations, and future research opportunities.

Skills and experience

The ideal student would have an interest in one or more of the following areas:

• Python programming.
• Traffic simulation or transport modelling.
• Artificial intelligence, machine learning, or large language models.
• Agent-based modelling.
• Human factors or driver behaviour.
• Automated vehicles and intelligent transport systems.
• Data analysis and visualisation.

Prior experience with SUMO or LLMs would be beneficial but is not essential. Prior experience with Python development is recommended

Start date

2 November, 2026

End date

19 February, 2027

Location

GP campus

Contact

Sebastien Glaser

3138 4911

sebastie.glaser@qut.edu.au