I wonder what the acceptable collisions/delivery needs to be for it to match last mile truck safety level (ie UPS trucks are big and run into things with non-zero frequency)

I'm sure there's a surprisingly high frequency of "acceptable" collisions if the bar is matching truck-inflicted property damage and injuries. Much like with replacing human drivers with computers, though, merely matching the cost and harms of the existing system is far from enough. Entrenched systems benefit from familiarity with the associated costs and risks, and from any structures built to mitigate them. New solutions have to be much better to gain acceptance.

Fortunately, automated systems can meet that higher threshold so long as we actually aim for it. If you aim for the lower "beats existing systems by some measures" bar then you make stupid decisions and tradeoffs like rushing to market or leaving out more capable sensors. We ought to try to make new technologies as good as possible. Sometimes the market will bet against that, but that's a tide that engineers should fight back against. Trucks kill too many people, and if drones kill half as many that's still unacceptable. We can do better.

People are more accepting when there's a person who can be punished. There's also the fact that society generally expects cars/trucks hitting things. A drone impact might be a more minor impact, but it's possible for it to hit things that are more shocking to the public if they get hit.