Allowing people to drive 2-tonne death traps around dense urban areas seems like a bigger mistake.
The fact that this prank can be done quickly, cheaply, safely, and without property damage means the car software needs to be prepared to deal with it without becoming a public nuisance. If the software is not prepared, the car is not fit for widespread deployment on public roads.
The equal possibility, which you raise, of dangerous impediments on both manned and unmanned vehicles is not relevant here, because the argument has not been made that the cars inadequately deal with dangerous impediments.
Finally, our society is largely comfortable with the current level of risk that manned vehicles entail, together with the ongoing progress of safety improvements. Not many would agree that the concept is a big mistake.
It might seem trivial, but it's not too hard to imagine a traffic cone on the windshield actually being a fairly dangerous situation if the car simply drove off, for instance if could roll over the top of the car and impact another vehicle behind it.
And that's even assuming that the traffic cone can be identified as 100% being a normal traffic cone, and not something else (say, a heavy object suspended in some meansm, in which case driving into it would damage the car) or a person or animal that happens to look like a traffic cone.
So by demonstrating that the cars stop under unexpected situations they are showing that the cars successfully identify a potentially dangerous situation and prioritize safety over navigation.
Also, I doubt it's a software problem: the cone is blocking a pretty substantial field of view. No amount of software can fix an occluded sensor. They might as well be spray painting over the cameras.