Closing the Modelling Gap: Transfer Learning from a Low-Fidelity Simulator for Autonomous Driving