Usage of End-to-End Machine Learning for Self-Driving Vehicle

Abstract: This thesis is about the usage of a machine learning model for self-driving vehicles running on a small mobile computing unit, in this case, a smartphone. We use NVIDIA’s PilotNet model which is a simple feed-forward machine learning model for steering a vehicle. Their testing is conducted on a real-world vehicle and the model is used for lane keeping. Instead, we’ve adopted to be able to drive around an oval track that has no lane marking. The primary goal was to be able to run the model on a smartphone and this was a simple task as we’ve seen that the inference time is small enough that it can run at 30 FPS. As of now the model only generate a good steering output in the testing phase with prerecorded data and was only able to complete a corner on one side of the track.