I don’t think stopping is a good idea in actual traffic.
Neither is running over a lump of clothes which may or may not contain a person. 
But humans are good at determining the difference between a lump of cloths and a person. Humans are also good at determining the difference between a scrap of cardboard and a brick wall.
What humans aren’t good at is determining the difference between a suicidal fox and an actual emergency. Instead of self driving cars, I would rather (everyone else had) a computer that stops people from doing stupid things. I can see the future now: a car that won’t let anyone lane change without signaling or shoulder checking; maybe we could use the airbag system as a tactile feedback system to alert the driver to their mistake.
So the ensuing car-viruses will actually drive me to Viagara stores and Nigeria?
I don’t see it happening anytime soon. The first car crash can make the car maker directly liable.
Human drivers keep liability to the owners, unless its a manufacturing defect.
A good comparison is subways and trains. Subways can be unmanned because the track is mostly secure from obstructions (but have a control room). Trains usually have a human monitoring it.
Roads are way more chaotic.
and brake for billboards.
But who is to say that the cardboard doesn’t have something behind it?
Honestly, the idea that a driver would just run into a cardboard box based on the assumption that it is empty is odd. If I see a refrigerator box on the highway, I sure as hell will not run into it thinking “La-de-daa, nothing will happen to me or my car 'cause I’m sure it’s empty.”
Just like seat belt alarms and seat load airbag arming systems now, make the auto drive refuse to engage unless there is a body in the seat.
You will still end up liable, just like if your dog bites someone. Auto driving cars are not going to run red lights or change lanes in the middle of an intersection.
I would tend to think even with some fairly chaotic road situations that computer driven cars will pretty quickly do alot better than human drivers, especially when a high percentage of the other cars on the road are auto drive as well…
They can see in all directions simultaneously
Computer reactions times are orders of magnitude faster than humans
Computers can know exactly how much braking force or accelleration to apply to not break traction, skid, fishtail, etc.
In the event of a collision that disables/disorients the driver but the car is still in motion the computers could still stop the vehicle safely and or avoid secondary impacts.
Other computerized cars around it will detect the evasive manuvers of cars around them in an accident situation and evade as well, minimizing multi car involvement.
Like trains these cars could sink up about a foot from each other gaining efficiency and always communicating with each other. And like birds in a field or a school of fish all emergency movements could be instantly communicated to adjoining cars. Plus when one car recognized the emergency it could communicate to cars behind and they could efficiently avoid it without slowing.
Major car manufacturers have been working towards this for decades. What makes Google think they’re going to get there first?
Google is doing the innovative stuff? Puhleeze. Google is great at self-promoting. I haven’t seen an actual car in operation yet and I won’t be holding my breath.
Google does not need to reinvent the car, the laser rangefinder, GPS, Radar, etc. They just need to build a software solution to make them all talk in a coherent fashion. Google does know a thing or two about software, probably a hell of a lot more than your average auto manufacturer.