Tesla's quarterly AP crash stats

[bhzmark]

Fixed it for you.

A little worrying that AP is trending towards more accidents / million miles.

Need to adjust for seasonality -- more accidents in winter. check year over year stats next year.

Although another difference will be younger less affluent Model 3 buyers which might make it trend down also.

stats are hard.

 

[lklundin]

To be clear, what I abbreviated as ”registered accident per million miles” it was meant to be short for registered ”accident per X million miles driven in which drivers had Autopilot engaged.” It was just shorthand.

I thought it would have been clear after the original post. It didn’t occur to me anyone could take the numbers the other way given the title and context given by the first post. I could have been clearer.

Is there a disagreement over the numbers?

Given your reading comprehension you should probably abstain from trying to read and redistribute non-trivial information.

 

[Eno Deb]

I hate to quibble details since I personally believe AP provides a huge safety improvement. But in terms of statistics, we should note that the places (e.g. highway) where AP is currently viable (and therefore mostly used) probably have fewer accidents per mile than other types of roadways (e.g. urban, city streets, w/ intersections). AP would probably show a better safety improvement in those areas, but we don't really know yet, because the data is under-represented.

Agreed. I wish they would either control for the driving environment of the non-AP miles, or drop publishing these misleading numbers. This isn't doing anyone any favors.

 

[J1mbo]

IMO a little more granularity would go a long way: road type, time of day, speed band; and maybe a controversial stat for the AP data to classify relevancy - AP is not fully at fault if the accident happens when the driver tries to use it around town and AP fails to stop at a roundabout, for example. Something fuzzy like "driver attentiveness ratio" or "operational domain ratio".

 

[electronblue]

Given your reading comprehension you should probably abstain from trying to read and redistribute non-trivial information.

Isn’t that a bit... unnecessary?

 

[electronblue]

Fixed it for you.

A little worrying that AP is trending towards more accidents / million miles.

Fair.

I had such a clear idea in my head of what we were talking about but it didn’t come out that way.

 

[ZeApelido]

Too many variables that aren't controlled here. One possible better method would be if Tesla accounted (say with GPS tracks) all the routes done on Autopilot, then look at non-autopilot safety numbers just on those same roads.

 

[electronblue]

Given your reading comprehension you should probably abstain from trying to read and redistribute non-trivial information.

Oh I see now.

You are TSLA stock gambler.

That explains why you blew a gasket on my completely innocent mistake.

Money...

 

[Ande]

Modern technology achievements described using ancient imperial mess.
Learn metric already ?

 

[NickFie]

Autopilot/Human (registered accident per million miles)

Q3/2018: 3.34/1.82 = 1.84
Q4/2018: 2.91/1.58 = 1.84
Q1/2019: 2.87/1.76 = 1.63

The ratios are interesting but not fully informative.

I see it as an answer to the question, "How does my accident risk per mile in Tesla using Autopilot compare with...

• Tesla not using Autopilot?
• NHTSA average?

The answer for the most recent quarter is: "On average, your Tesla with Autopilot risk per mile is...

• 46% lower than Tesla without Autopilot. ( 1.76 / 2.87 ) - 1 = -0.46
• 85% lower than NHTSA average. ( 0.436 / 2.87 ) - 1 = -0.85

Interesting that Autopilot miles/accident is trending down. I attribute that to the growing proportion of novice Autopilot users due to Model 3 sales surge.

Who's worried? Son-in-law sets out on nearly 400-mile journey, with his daughter, in our S 100D tomorrow morning. He's driven the S frequently, including a road trip. Still, this will be his first trip using Autopilot and NoA.