End-to-End Machine Learning for Autonomous Racing
Project Overview
Teaching a neural network how to drive on racetracks it has never seen before.
What were the key results of your research project?
The central result was evidence of generalisation: a neural network trained only on the Budapest circuit was able to drive Monza, a track it had never encountered with no retraining. It reached the same maximum stable speed on both tracks (5.0 m/s), which suggests the network learned a transferable driving skill from the laser sensor rather than simply memorising one specific track layout.
I built a complete working pipeline end to end: recording driving demonstrations in simulation, converting them into training data, training the network, and deploying it to drive in real time at 40 inferences per second. Across the project this generated roughly 532,000 training examples.
I also upgraded the "teacher" that generates the training data. The initial expert algorithm was cautious and imperfect, so I built a second one from scratch that follows an optimal racing line and sets an appropriate speed for each corner producing more, and cleaner, demonstrations for the network to learn from.
Honest caveats: this was an 8-week project, so the generalisation result rests on one unseen track rather than many; everything was tested in simulation, not yet on the physical car; and the improved teacher needed per-track tuning that didn't always work reliably.
A presentation of this research will be shared here and on our YouTube site when available.
How do you feel you have benefitted from completing this internship and has it made you consider future career paths?
This internship has given me a much clearer understanding of what working in a professional research environment actually involves. Beyond the technical work, I learned how to scope a project realistically, make decisions about direction, and communicate progress through regular meetings with my supervisor, experience that's quite different from typical coursework.
On the technical side, I developed a far deeper understanding of ROS2 and how the different components of a robotics system fit together, from sensor data and message passing to real-time control and deployment. Setting up and debugging these systems myself, rather than working from a tutorial, gave me a genuine working knowledge of the full stack I was using.
The project also shaped my sense of direction. I found myself increasingly drawn to the machine learning side of the work such as designing the model, understanding how it learned, and testing whether it could generalise. This has made me seriously consider postgraduate research, as I enjoyed the open-ended, investigative nature of the project and would like to explore machine learning in more depth.