perception, planning, prediction, eval
In a pipeline, the fragile place is often not inside a station. It is between stations.
In the children's game of telephone, each person hears clearly and speaks clearly. Nobody makes an obvious mistake. The sentence still arrives wrong, because every handoff loses a little and the losses stack.
Separate driving modules have the same failure mode. Perception summarizes the world into a tidy list of objects with positions and speeds. That list is handed forward. Summarizing means discarding, and the discarded detail sometimes mattered.
Nobody makes an obvious mistake.
Example: perception reports a pedestrian's position, size, and heading. It may not pass along that the person was looking at a phone, had one foot angled toward the street, or was carrying something that could drop. Prediction can only work with what it received.
Confidence is the second leak. Internally, perception may be only partly sure an object is a cyclist rather than a sign. If the interface only passes the word cyclist, later stages plan as if the label were certain. Uncertainty disappears exactly where it was useful.
That is one reason many teams train stages together. A joint representation does not have to be a human-readable object list. It can carry whatever turned out to help, including nuance nobody wrote down in a spec.
Joint training has costs too. Those costs are the next posts.