This article makes the assumption that we are learning a complete model that goes from sensor inputs to control outputs, but I don't think anyone is doing this

This article makes the assumption that we are learning a complete model that goes from sensor inputs to control outputs, but I don't think anyone is doing this outside academia. There's a whole lot less controversy when we use deep learning to do scene understanding, where we understand at a high level the model is recognizing entities in its sensors, and we can evaluate whether that subsystem failed, etc.