training, fleet, eval
The first 95 percent of driving is repetitive. The last 5 percent is almost the entire project.
A playlist of the same twelve songs feels finished after a weekend. A playlist that must include every song anyone might request is never finished. Ordinary highway miles are the twelve songs. The long tail is the request line.
Mattress in a lane. A person in a duck costume. A police officer waving through a red. A flock of birds that lift late. A moving truck blocking both a stop line and a crosswalk. Each event is rare in any one driver’s life and common across a national fleet.
A playlist that must include every song anyone might request is never finished.
Rule-based stacks drowned here because each rarity wanted a new clause. Learned stacks drown here in a quieter way: if the training set barely contains the event, the model has no strong example to copy. Collecting the rare thing becomes the job.
That is why fleet scale shows up in every serious conversation. You do not need a billion boring miles to learn a left-turn-on-green. You need enough of the ugly leftovers that a person only sees a few times a decade. Most of the remaining work in autonomy is search: find those leftovers, train on them, check that the fix did not spoil the twelve songs.