I'm inclined to agree, especially because it helps in 1) providing diagnostic information (such as the great driving visualizations shown in the video), and 2) makes it easier to incorporate algorithms and sensors (like with Google's cars) as a redundancy in case the neural network hits a crazy edge case.