This paper proposes a strategy for collecting road frames and steering data for training deep end-to-end control policies for robust lateral vehicle control. During the data collection phase, one camera is deployed in the vehicle and two steering controllers are operated simultaneously. While one controller operates in closed-loop and keeps the vehicle on a desired perturbed trajectory, a second controller computes the nominal steering wheel angle required to drive the vehicle to the center of the desired lane. With this approach, it is possible to train a convolutional neural network to act as an end-to-end lateral vehicle controller that is capable of rejecting unforeseen disturbances. We implement our approach by incorporating the deep learning framework Caffe and the vehicle simulation software TORCS in Matlab/Simulink and analyze the robustness of the trained end-to-end control policy in closed-loop simulations.

This content is only available via PDF.
You do not currently have access to this content.