Project Abstract

This project explores the application of reinforcement learning (RL) to autonomous driving within the realistic racing simulator Assetto Corsa, using only raw screen input and keystroke-based control. With the increasing relevance of RL in real-world autonomous systems and gaming, this work investigates whether a deep RL agent can learn to navigate complex tracks using visual data alone, without access to game telemetry or internal state information. The project aims to build an end-to-end pipeline that captures screen frames, preprocess them and then feed them through a convolutional neural network (CNN), and uses a Deep Q-Network (DQN) to select from eight discrete driving actions, each mapped to specific keystrokes that control the car’s steering, acceleration, and braking. DQN was chosen for its simplicity and proven track record in vision-based RL tasks such as Atari games, while the reward function was carefully shaped to encourage lap completion, staying on track, and penalizing crashes or erratic behavior. The trained agent demonstrated an ability to complete laps under certain conditions and learned basic driving behavior such as lane-following and throttle control. However, performance dropped in areas requiring fine control, such as sharp turns, due to limitations in the discrete action space and sparse reward signals. This work provides a practical example of RL applied in a more realistic, modifiable simulation environment and highlights key challenges in action discretization, real-time decision-making, and training stability. It contributes insights into how vision-based RL agents can be developed using accessible tools and games, offering a foundation for future work involving more advanced algorithms, continuous control, and domain transfer to real-world scenarios or physical autonomous systems.

Keywords: Reinforcement Learning, Racing Simulation,

 Conference Details

Session: A

Location: Sir Stanley Clarke Auditorium at 11:00 13:00

Markers: Matt Roach, Jen Pearson

Course: MSci Computer Science 4yr FT

Future Plans: I’m continuing studies