There's a reason that there's no path to doing this legally.
1) Have you ever flown an RC plane? Do you know a) How difficult it is? b) How little control you have? c) How dangerous it is if not controlled? That's why (sane) people fly their planes in big fields away from people, trees, power lines, clotheslines, satellite dishes, etc.
2) More broadly, we are only beginning to have good algorithms for awareness of automobile traffic conditions. (Google's work with driverless cars, for example.) This is basically 2-D work on roads, which are specifically designed to avoid interference from external sources. None of these apply to drone aircraft. In addition to the obstacles above, how do avoid all the potential objects (other drones, birds, party balloons, etc) in proximity? We're nowhere near close to solving these problems.
3) Yes, there are accidents by delivery drivers every day. But it's clear who's responsible for any damage: the human driver(s) who are presumed to be carefully monitoring the entire process. Are mom and pop taco houses (or any other sized business) prepared to take the full liability of an autonomous flying thing? I doubt it.
All the videos we've seen (and I admit some are breathtaking) have been created by people firmly at the controls, whose full attention is focused on the task, and who are prepared to take the responsibility for any problems they cause.
Quadrotor robot control is pretty good already and getting better quickly. You can buy these things off the shelf with GPS waypoint navigation. Completely autonomous. Not perfectly reliable in all conditions, but already usable on a calm day. Thus the control difficulty is not an issue in the medium term.
Many people have enjoyed the quadrotor videos from labs such as UPenn's GRASP lab. No one at the controls. Though the miniature flyers at Penn do use external localization and pre-computed trajectories, others are completely autonomous and react in real time.