Biography

Constant Bridon works as a data scientist at OCTO Technology.

Talk

Racing an Autonous Toy Car from Scratch Powered by Machine Learning

A racing competition held in France asked players to design an autonomous toy car, piloted with deep learning models. Following a line is a well known problem in computer science, and there is no need for a neural network to perform such a task. However, I decided to assemble hardware components (rapsberry, arduino) and design some code in python to make an autonomous radio control car, using only a camera.

The deep learning approach implies to have training examples to quantify the weighs of the architecture. You can either have real case examples that you generate while manually piloting the car. This method can be defined as shaping the model, which is not satisfying because it may produce unexpected biaises. I decided to adopt a simulation technique to generate my training examples. As a result, the car is autonomous without having ever driven on the race. I would like to share with you the general workflow, from scratch, ranging from hardware assembling to racing in production, via training models.