Autopilot data growing 1.5M miles/day. Tesla knows by now if it's safer than a human

[JeffS]

tmoz - I think you need a service visit and take a tech for a test drive to replicate the situations. At a very minimum, it would seem your camera needs to be calibrated. My car does not bias right. And a friend's car does not bias right. Especially in the left lane. Something's not right with your system. Until that time, with the situations you described, I can totally see how it is not smart for you to use AP at all. Not even a little bit. Especially the part about aborting AP and you not realizing it through the visual and audible cues, and then drifting into another lane. If an accident occurred in that situation, it would be very sad.

 

[Marc_GA]

That 6.1 accident per million miles is an overall average, and therefore only valid for a typical driving profile. Since AP miles are mostly freeway or country road miles, I don't think it's reasonable to expect 6.1 per million - those are generally much less likely places to have an accident than city streets or neighborhoods.

Using the NY DOT numbers (https://www.dot.ny.gov/divisions/operating/osss/highway-repository/Table2_2014.pdf) it looks to be between 1.01 accidents/million miles (Controlled Access, excluding intersections/ramps) to 2.44 acc/MM (Controlled Access, undivided rural). Going to less controlled roads (expressway, highways) raises that number up to 4.41 worst case (acc/MM).

Personally, we have done 24k miles in a year, unfortunately most before AP existed. Since then 3k of the 5k or so have been AP enabled.

Just taking those at face value would have 1 accident per day (assuming the 1.5MM/day is correct). I personally think this is a little high but for a Fermi estimate it's in the right ballpark.

 

[Saghost]

Using the NY DOT numbers (https://www.dot.ny.gov/divisions/operating/osss/highway-repository/Table2_2014.pdf) it looks to be between 1.01 accidents/million miles (Controlled Access, excluding intersections/ramps) to 2.44 acc/MM (Controlled Access, undivided rural). Going to less controlled roads (expressway, highways) raises that number up to 4.41 worst case (acc/MM).

Personally, we have done 24k miles in a year, unfortunately most before AP existed. Since then 3k of the 5k or so have been AP enabled.

Just taking those at face value would have 1 accident per day (assuming the 1.5MM/day is correct). I personally think this is a little high but for a Fermi estimate it's in the right ballpark.

Seems reasonable. It'd be interesting to see what Tesla's data is. Of course, if Autopilot stops you at the back end of a traffic jam and some idiot who's busy texting rear ends you, that would still show up even though it isn't AP's fault - you'd need some way to account for accidents that aren't the system's fault.
Walter

 

[tmoz]

Yes, I'm pretty nervous with AP on, not relaxing. I did have the rear sensors replaced when I first got the car as they were faulty. I'm 5 months away from the annual inspection, that is probably too long to wait. I should talk to the guys and see if they can check it out.

tmoz - I think you need a service visit and take a tech for a test drive to replicate the situations. At a very minimum, it would seem your camera needs to be calibrated. My car does not bias right. And a friend's car does not bias right. Especially in the left lane. Something's not right with your system. Until that time, with the situations you described, I can totally see how it is not smart for you to use AP at all. Not even a little bit. Especially the part about aborting AP and you not realizing it through the visual and audible cues, and then drifting into another lane. If an accident occurred in that situation, it would be very sad.

 

[David_Cary]

Perhaps not entirely fair to throw this in there but....

Accident rates are very dependent on things you wouldn't necessarily expect - like education level, and those you would like age. I looked into this with all the fatality discussion. If you take a middle aged 100k income, some post college - your fatality rate is tiny and presumably so is your accident rate. So comparing to averages just isn't reasonable at all.

So at some point, AP can be better than average driver (maybe not hard) but not nearly as good as average Tesla driver.

The education difference for fatality is 4.3 to 1 (actually looks like more like 8 to 1 for grad school to less than high school but 95% CI probably brings it to 4.3).

Now suicide is another manner - based on epidemiology and actual Tesla history.

 

[mgboyes]

Surely Tesla simply need to compare the accident rate within the Model S fleet before and after Autopilot was enabled? That way all other factors will cancel out. Who cares what the national average accident rate is - the only reasonable comparison is to compare Tesla with Tesla.

What I don't know is whether Tesla have sufficiently well-structured data that they can say with confidence how many accidents Model S's had while in Autosteer-suitable situations, before Autosteer was released.

 

[eloder]

The data would be biased. People only engage AP when they think it would be safe to do so.
So the 1.5 M miles per day tells us how safe AP is when it is allowed to operate in situations that the drive deems safe.
Extrapolating past that point would not be justified.

True, but the thing you must realize is that humans are not immune to accidents when driving in conditions that they believe safe to do so. If AP data shows statistically significant decreases in--say--straight-line highway driving, that's still a huge win.