End-to-End Learning of Autonomous Vehicle Lateral Control via MPC Training

One of the main requirements of an autonomous vehicle is the ability to maintain its trajectory within the road lane. This task is generally performed utilizing vision data, processed using convolutional neural networks or classical computer vision algorithms to extract a road mask. A software pipeline then analyzes this mask to retrieve the vehicle's relative state. This process is composed of many components that need to be tuned to achieve good results. What is proposed in this paper is instead an end-to-end solution able to infer the steering command directly from camera images. Differently from the classical end-to-end machine-learning approaches, the architecture is not trained using as ground truth the car data from a human driver, but instead the output of a control algorithm. The network then does not mimic a specific human behavior but learns how to achieve the optimal trajectory computed by the algorithm in an end-to-end fashion.