The problem is: real-life driving does not only consist of situations that even a child could easily manage. Humans are extremely good at letting their driving skills degrade gracefully when conditions get rough. AIs? They drive perfectly until they reach their limit, then they suddenly pull off epic fails - if no human driver is on standby to resolve the situation immediately.
And programmed to turn off the blinker light after merging onto the freeway...
They totally drive by night too. How do I know that? Simple: You see them with your own eyes, in Mountain View. Even if Google doesn't brag about it.
> Mastering autonomous driving on city streets -- rather than freeways, interstates or highways -- requires us to navigate complex road environments [...]. This differs from the driving undertaken by an average American driver who will spend a larger proportion of their driving miles on less complex roads such as freeways. Not surprisingly, 89 percent of our reportable disengagements have occurred in this complex street environment (see Table 6 below).
1: https://static.googleusercontent.com/media/www.google.com/en...
Statistically, Google's Self-Driving Car would've lost it's license by now, if not for human drivers keeping it in check.
PS: How often do you think the average drivers ed teacher uses there break?
A minor fender bender at low speed vs. losing control and going over a guard rail at 80 mph are both "accidents" - but have entirely different consequences.
I have a hunch that while human drivers are still better, when they do have accidents some percentage of those tend to be much more fatal.
424,331 miles, according to the report. http://www.consumerwatchdog.org/resources/cadmvdisengagerepo...
Nor are you counting all of the times the driving software handed control back to the driver. Consider each one of those equivalent to your human driver falling asleep at the wheel.
Our objective is not to minimize disengagements; rather, it is to gather, while operating safely, as much data as possible to enable us to improve our self-driving system. Therefore, we set disengagement thresholds conservatively, and each is carefully recorded.
That's a long way from falling asleep at the wheel. That's we don't want to put people at risk so we are going to hand off control if "Disengage for a recklessly behaving agent" etc.
But the report actually goes so far to points out that on at least 13 of those occasions Google believes that without human intervention a crash would have occurred. It also implies that only 3 of those situations were created by poor driving on the part of another human. That's a definite crash due to Google car error averted every ~40k miles, even though humans already take over whenever there's a software warning, a spot of rain or something else that they think might stretch its capabilities too much.
The average US driver has 1 accident per 165000 miles (which, like the unusual tendency for the Google car to get bumped when stopping at junctions, may or may not be their fault) and that's including DUI as well as people driving in slightly more difficult conditions than sunny surburban California.
The average driver doesn't do all city miles and that's why Google's cars have gotten bumped. Highway driving is much easier to automate (several cars already do this today) and makes up a large percentage of total miles driven.
That is a useful data point. Extremely optimistic bystanders think self driving cars will lower death rates to zero.
Statistically about 7 people die on the public roads per billion passenger miles. At half a million miles, assuming the miles are statistically random and indistinguishable from the entire country (sure, summer in socal is as hard driving as a winter in a blizzard at night in Montana, sure...) then 424331 death free miles means that technologies death record is no worse than 336 times worse than human drivers, at least so far. Perhaps its only 335 times more deadly than human drivers. Even drunk people aren't that dangerous.
Or in other words, the self driving car can be summarized to lots of predictions based on very little data.