fleet, training, eval
A million copies of the same commute are one lesson, repeated loudly.
Doing the same gym circuit every day for a year does not make you ready for an unplanned sport. It makes you very good at that circuit. Fleet miles work the same way.
If most recorded miles are dry highway following, the network gets a loud signal about dry highway following. Rare city construction stays quiet. Adding more of the loud miles can even bury the quiet ones, because training spends most of its attention where the data is dense.
It makes you very good at that circuit.
What matters is coverage, not a raw odometer. Teams talk about unique scenes, interventions, and mined clips because those are ways of counting lessons instead of minutes. A short clip of a rolling ball at a curb can outweigh an hour of cruise.
This is the correction to a popular headline number. Miles are easy to advertise. Diversity is harder to measure and easier to fake with sloppy sampling. A smaller, better-searched set can beat a larger, repetitive set. Scale helps only after you can find the tail.