planning, training
For years, a lot of driving software was a recipe a person sat down and wrote.
A written recipe is useful. Someone else can follow it. If the dish fails, you can often find the step. A recipe only covers what the author anticipated. It does not tell you what to do when the cream splits, the oven runs hot, or an ingredient is missing. An experienced cook handles those without opening the card.
Early stacks looked like recipes. Engineers wrote conditions: if the signal is red, stop. If an object is in-lane within this distance and closing at this rate, brake at this deceleration. If the lane marking curves past this threshold, follow it with this steering request. The behavior lived in large trees of human-written cases.
Someone else can follow it.
That design had real strengths. Behavior was inspectable. A mistake could often be traced to a rule. Some properties could be argued from the code. One fix could be isolated from another.
Coverage was the ceiling. Streets produce situations nobody listed — a mattress in a fast lane, a parade, a person waving traffic through a red light, birds on the asphalt. Each new case wanted a new rule. Rules interacted. The system could get more complete and more brittle at the same time.
At some point teams decided the recipe would not scale. What they replaced it with is the next post.