Autopilot Safety: 2-3x safer than average really safer?

[tander]

So I listened to a recent Musk quote about self-driving (

) and he said that Autopilot should be 2-3x safer than the avg driver. 2-3x safer sounds pretty great, sign me up. But I'm wondering if that's really that good? The reason is if most people are good drivers that rarely get into accidents, and most accidents caused by a very small amount of bad drivers that get in lots of accidents, then most drivers are much better than the average right? So is 2-3x safer than average actually any safer than a non accident prone driver? Anybody that's waded into the stats car to opine? Maybe I'm making assumptions or misunderstanding?

 

[Derek Kessler]

That's not how averages work. When you have a large, generally homogeneous sample the outliers have little impact overall.

Say you have a sample of 100 drivers, and we can rate them on a 1-10 scale of driving proficiency. If 95 of our drivers are excellent 9-point drivers and 5 are terrible 2-pointers, our average comes out to 8.65 — not far off from the 9.0 all-star drivers, even with those five awful jerks wreaking havoc on the highway.

What Elon is saying is that a self-driving car should score something like 20+ points on the 10-point scale because it'd be that much better than we dumb humans. We're still a long way off from that, though Tesla does claim that Autopilot-engaged cars are statistically something like 50% safer (so... 13/10) than humans alone.

 

[Az_Rael]

I guess the problem is everyone thinks they are an above average driver. But that can’t actually be true. I'm the good driver, you're the bad driver - CNN

I know I think I am above average, but I sure do occasionally make a boneheaded move while navigating the roads. Allstate says here https://www.allstate.com/tools-and-resources/americas-best-drivers.aspx the average driver experiences a collision every 10 years, so I guess I am not beating that metric (been hit twice since moving here in 2010)

So, where do I fall on the actual driver skill scale? I don’t know. I hope that 2-3x safer means 2-3x safer than ME, not 2-3x safer than a driver worse than me. But I do contemplate that whenever I engage AP2, and it does something wonky so I have to turn it back off.

Here is at least one report:

U.S. Drivers Score an Average "C" Rating on Safe Driving Skills

I would guess the folks who know best what average driving actually is are probably insurance companies. They do all that data scrubbing to basically profile folks into risk categories.

 

[tander]

@Derek Kessler that's a good example, it makes sense. Maybe I'm overthinking it but thinking about it in terms of @Az_Rael, I wonder if the statistics/reality are not as simple as 1-10 scale, I guess I might be able to answer that if I was an actuary. For example if the average driver has a collision every ten years, well I haven't had one in 20 years (knock on wood, and yes I do boneheaded things sometimes too) so should I get a score of 20? How do you define good driver anyway does fault matter or just whether or not you have an insurance claim? There's probably a lot of good drivers that unluckily get rear ended so to me that doesn't make them a bad driver. Should a driver that causes a wreck get a negative score? What about drivers that prevent what would otherwise be a wreck? I guess what I'm really wondering is how nuanced/accurate that comment was from Musk, and if Autopilot will continue improving until it's 10x safer or what, because eventually, hopefully soon, it gets to the point where it's just so good that it's irresponsible to let people keep driving and all this is moot I guess. According to google there are 6 million accidents a year in the US so I guess if you could knock that down to 2-3 million that's a big win even if the goal is zero.

 

[outie]

I know with AP engaged it's tried to kill me and my family several times by running into the divider or adjacent cars on the freeway. However, because of knowing something is trying to kill me I am staying quite alerted so yes he's right it's 2x-3x safer as we have AP doing the driving and me watching over it.

 

[ShockOnT]

The real problem is that people will have a hard time accepting thousands of deaths caused by autonomous vehicles, even if it replaces tens of thousands of human-caused deaths.

 

[theflyer]

The real problem is that people will have a hard time accepting thousands of deaths caused by autonomous vehicles, even if it replaces tens of thousands of human-caused deaths.

This will be true initially. You need to look no further than the death in Tempe. However, I would posit that because autonomous auto deaths will be of limited scale per event (e.g., 1-5 people), society will eventually move past the hype and deaths will blend into the background just as they do today with human drivers. As a species, we are abysmal at internalizing scale when it happens over long periods of time or in small numbers per batch.

 

[WannabeOwner]

Problem I have with this is that because I'm a good driver maybe I could avoid the Machine's accident ...

... and if that accident would kill the driver, not the other party, then I'm not happy with that. So I would want the ability to override ...

... will other people want that too, or will they be happy with "the odds" ?

I like AP because I don;t care if AP or ME detects a risk, operating together we are better than ME alone. Particularly when I'm eyes-off-the-road, momentarily, or car two-in-front brakes but I cannot see it.

 

[strangecosmos]

I really want to know how much safer than average the top percentile of drivers are. If they’re only 2x safer than average, okay, no problem. But if they’re 5x safer than average, then they’ll actually be less safe getting into a self-driving car that’s only 2x safer than average.

The top percentile should also be a benchmark, not just the average.

 

[Snerruc]

This problem is more complicated than being discussed. Many years ago I read a study of the bad driver/good driver scenario. As best as I can recall, by statistical measures, good drivers were more than one order of magnitude better than bad drivers. One complication cited in the study was that there are many drivers that are “bad” who have excellent records. Myself, for example. For two years I was blind and, when my sight was restored, , it was far from perfect. I accepted that this made me a bad driver as I often don’t see things properly. I thought myself compensation techniques with the result I have not had an accident or ticket since ( well over 20 years). There are also drivers with excellent skills that are dangerous drivers. If Autopilot took me off the road, it wouldn’t accomplish much. If it took that “good”driver off the road, it might.