I've never been convinced that deep learning is the correct approach for something so nuanced as driving a car. Deep learning is excellent for generalisations and consolidating data into single metrics. It's not so good at identifying edge-cases and reacting accordingly, those cases are more likely to be merged into other reaction types.