AI-Based Driver Behaviour Simulation

The problem

Driver training needs vehicles, instructors and controlled road sessions, which is costly and exposes trainees and other road users to risk. Simulation can repeat the same situation safely. The deeper modelling question was how a simulation can represent realistic human decision-making instead of assuming every driver follows the optimal rule.

Research question

What action is a driver likely to take, given the environment, their risk tolerance and their reaction time, rather than what is the safest possible action?

System design

Scenario world → simulated sensors (with noise and a reaction delay) → feature extraction → decision tree (scikit-learn, trained in Python, run in the browser) → rule and safety layer → control (acceleration, braking, steering) → vehicle movement → logged metrics. Driver profiles change risk tolerance, reaction time, target speed and following distance.

Method

The original slides describe three stages: object detection, decision-making and interaction implementation. The 2024 prototype was built by modelling a road, junction and traffic light in Blender, exporting to Unreal Engine 5, setting up a traffic-light Blueprint (red, green, yellow with delays), collecting 12,758 pictures and integrating an AI-driven car. The recreation expands the same idea into simulated sensors, feature extraction, risk estimation, a decision tree, a rule-based safety layer and vehicle control, driving a route planned between two places on a map.

Original: the slides conclude that the small prototype detected obstacles and made decisions, supporting the feasibility of AI-driven driver-training simulation. Recreation: a 2-D top-view map where you choose a start and a destination, plus seven controlled test scenarios, run with three driver profiles or your own settings, with a live decision panel, event log and metrics, a profile comparison and a failure matrix computed in your browser from seeded runs. Numbers come from running it, not from this text.

What it does not show

The original Blender and Unreal source files are not available, so the recreation is new and simplified, not a port. The map and the place names are a provisional layout traced from a slide until the real map is supplied. The model is a decision tree trained on synthetic scenarios labelled by an authored braking policy, with no real driving data, so its accuracy measures imitation of that policy. Sign and hazard detection are simulated from scenario data and are not computer vision. Driver-profile parameters are illustrative.

Lessons

Modelling realistic rather than optimal behaviour meant deliberately encoding reaction delay and risk tolerance instead of treating deviation from the rules as pure error, and keeping that variation bounded so it stays tied to the situation. A learned model also needs a rule layer around it and an honest account of what it was trained on.

Responsible AI

This prototype is for research, demonstration and simulation. It is not an autonomous-driving safety system and must not be used to control a real vehicle. Near misses and contacts are shown as data to analyse, not as a game.