Once you start optimizing across large numbers of drivers you end up dispatching the same driver to the same restaurant to do multiple pickups anyway, which then need to be delivered quickly. Outside of peak times having fewer drivers than restaurants seems like a net win, but restaurants can make use of drivers when their not making deliveries. This also means their in the restaurant and thus lowers delivery times.

Believe me I understand why it seems possible, but this is one of those cases where real world data doesn’t fit abstract models very well. One example is during peak times both the restaurant and it’s drivers get overwhelmed so preparation time is increasing. Another example is if their returning to the same location they can easily return empty insulated boxes, but that’s problematic if their doing pickup from 10+ restaurants a shift.