A neural network-based controller is trained on data collected from connected human-driven
vehicles in order to steer a connected automated vehicle on multi-lane roads. The
obtained controller is evaluated using model-based simulations and its performance
is compared to that of a traditional nonlinear feedback controller. The comparison
of the control laws obtained by the two different approaches provides information
about the naturalistic nonlinearities in human steering, and this can benefit the
controller development of automated vehicles. The effects of time delay emerging from
vehicle-to-everything (V2X) communication, computation, and actuation are also highlighted.
