Seems like Google acknowledged this:
"From now on, our cars will more deeply understand that buses (and other large vehicles) are less likely to yield to us than other types of vehicles"
"From now on, our cars will more deeply understand that buses (and other large vehicles) are less likely to yield to us than other types of vehicles"
It will be an endless process for an algorithm driven car. Buses in Dublin will be so much different from buses in London, or trucks in the US different from trucks in Germany. Then there are the bikes, pedestrians. When human drivers/bikers/pedestrians look at each other and know what to expect. These are the things that would be incredibly difficult to formalize, if possible at all. Unless you have something like a Deep Mind on board, but even then I wouldn't be so sure.
This is already an issue humans solve in the context of tourists driving through town, who aren't accustom to local rules like "First driver at an intersection gets to make a left on green instead of yielding to the column of oncoming straight traffic."
Signalling intent to change lanes indicates that you are weak; people will actively cut you off.
Even if they're not perfect, they're still already better than the median driver in the majority of daily driving. They don't need to be perfect, just better than most and you're drastically reducing deaths. Below median drivers cause more accidents than above median.
Most accidents come down to awareness and reaction (both reaction speed and the ability to evaluate the right course of action). These are two things that a sufficiently advanced computer system will always be better at. A human gets distracted, doesn't have 360-degree constant vision, doesn't have thermal imaging or the ability to range-find obstacles through fog. A human can't use perfect situational awareness of obstacles and road conditions to avoid an accident.
Anyway, the plural of anecdote isn't data. Google is in the data business and has a firm grasp of what they're attempting. I doubt they're going to be proven wrong.
From a utilitarian perspective, they would only need to be better than whoever they are replacing one driver at a time.
From a human-feely point of view, they'll need to be massively better than the best human to stand a chance of adoption.
My guess about AI: for human-suitable tasks the gap between a typical human and the best human is actually not all that great, compared to the difficulty of getting to human-level skill in the first place. So by waiting for surpassing the best human, we don't actually lose too much time.
Artificial Neural Networks are actually totally deterministic by default. (Even if training them is not.) What they lack is people being able to explain in simple ways what the ANN is doing.