Hand coding the output of perception systems for self-driving cars is insufficient because it will miss important features in the real world that are not present in simulated environments, making the feature vector look fundamentally different.