After a long year, I am happy to share that I have finished my Master’s dissertation titled Adaptive Motion Control for Autonomous Racing. I managed to create a 2-stage system consisting of trajectory optimization and tracking, both of which share the same model and adapt to dynamical changes online. The key is that the model is a semi-parametric one consisting of a classical parametric model which does the larger part of the prediction and a neural network which learns the residuals (or errors) of the parametric model. This method, combines the strength of both modelling types, resulting in a system which generalises well, learns accurate dynamics and adapts robustly to changing dynamics. Sadly, it can only adapt to short-term changes in dynamics due to the batch training nature of neural networks.
Even though I didn’t achieve everything I had planned to, this experience has taught me to look past the idealistic aspects of both motion control and machine learning, but instead understand their fundamentals. I now realize that is more important than getting good results.
A big thank you to Michael Mistry, Christoforos Chatzikomis, and Timo Völkl for supporting me through this journey. I wouldn’t have done it without you!
Sadly I cannot share my dissertation publicly, but you can find part 1 of my dissertaion which sets the grounds here