Tesla Has Accumulated An Incredible Amount Of Autopilot Miles


JUL 17 2018 BY EVANNEX 27


The MIT Autonomous Vehicle Technology Study (MIT-AVT), which is studying human-AI interaction in driving, has collected masses of real-world driving data and is using that data to glean all sorts of insights of interest to Tesla fans. In a previous post, we admired a set of nifty charts, prepared by study co-author Lex Fridman, that illustrate how many Tesla vehicles with each version of Autopilot hardware are on the roads.

*This article comes to us courtesy of EVANNEX (which also makes aftermarket Tesla accessories). Authored by Charles Morris. The opinions expressed in these articles are not necessarily our own at InsideEVs.

Above: Tesla Model S (Image: Tesla)

Now Mr. Fridman has aggregated various publicly available data sources to elucidate the total number of miles driven by Autopilot-equipped Tesla vehicles.

Fridman has charted the number of miles driven using both Autopilot hardware versions, as well as the number of miles driven in shadow mode (in which Autopilot logs the data streams from vehicles even when they are under manual control).

Above: Estimated miles all Tesla vehicles have driven including those in Autopilot “shadow” mode (Source: MIT-AVT)

Fridman estimates that Tesla vehicles have driven over 7.8 billion miles in the aggregate, over 1.2 billion of those miles in Autopilot mode and over 1.6 billion in shadow mode. Even more impressive than the size of these figures is how quickly they were reached – less than 3 years. With deliveries accelerating, the number of miles, and the amount of data gathered will soon go into hockey-stick mode, keeping Tesla miles ahead of any potential competitors in the vehicle autonomy race.

According to Fridman’s calculations, the average number of miles driven per Tesla vehicle per day is 31.76, and the average number of Autopilot miles driven per Autopilot-enabled vehicle per day is 7.91. This means that Autopilot-capable vehicles are operating in Autopilot mode about 25% of the time.

Above: Estimated Autopilot miles driven in both Tesla’s hardware 1 and 2+ modes (Source: MIT-AVT)

As Fridman sees it, Tesla is striving to develop AI systems that save human lives on a very large scale. “The stakes are high and the pressure on engineers to do the best work of their life couldn’t be higher. We have a lot of data in the MIT-AVT study that helps illuminate how to take on this life-critical challenge.”

Fridman and his colleagues invite all Tesla owners to help with the research by taking (and sharing) his team’s detailed survey on Autopilot.


Written by: Charles Morris; Source: MIT Autonomous Vehicle Technology Study

*Editor’s Note: EVANNEX, which also sells aftermarket gear for Teslas, has kindly allowed us to share some of its content with our readers, free of charge. Our thanks go out to EVANNEX. Check out the site here.

Categories: Tesla


Leave a Reply

27 Comments on "Tesla Has Accumulated An Incredible Amount Of Autopilot Miles"

newest oldest most voted

Correct me if I’m wrong here, but there’s a fatal crash in the US every 130M miles, right?

So with 1.2B autopilot miles, there should have been 9 fatal crashes by now. Instead, there’s only been 2 or 3. So we can say someone in a Tesla with autopilot turned on is at least 3x safer than the average person in a car in the US.

Can’t conclude that that is entirely from autopilot, or even that autopilot contributes to it – too many factors involved. But still an interesting stat. I suppose if someone knows how many fatalities have occurred in all Teslas combined as opposed to only ones with autopilot engaged, we could remove several variables and zoom in on how many lives are saved (or lost) by autopilot software as opposed to by the physical structure of the cars.

Good points, but that would be a bit difficult to calculate in the comments section. That is the size of study that PhD students get funded to do….

It’s too early to say, but there have certainly been more 3 fatalities with autopilot turned on. There may have been only 3 that you have heard about. Regardless, I think eventually it will turn out that you are correct and one will be 3-5x times less likely to die in a Tesla that crashes with autopilot engaged, than the statistical norm for fatalities resulting from accidents. Autopilot continues to improve.

“So we can say someone in a Tesla with autopilot turned on is at least 3x safer than the average person in a car in the US.”

1. Quibble: It’s Autopilot+AutoSteer, not Autopilot alone. Autopilot is always on, and can’t be turned off, while the driver can turn AutoSteer on and off. But Tesla’s PR also says “Autopilot” when it means “AutoSteer”, so you’re in good company. 😉

2. Only 2 or 3 fatalities in Tesla cars under the control of Autopilot+AutoSteer is much too low a statistical sample to have much meaning. Far better to use the overall accident rate. The NHTSA says that Tesla cars with Autopilot+AutoSteer merely installed — not necessarily active — have a 40% lower accident rate, as measured by airbag deployment. (That is, serious accidents, not minor fender-benders.) So it’s pretty straightforward and simple logic to conclude that with AutoSteer active, the accident rate must be lowered by significantly more than 40%.

It’s amazing how hard the Usual Suspects (among the serial Tesla bashers) try to deny these solidly established facts!

Um… What exactly is Autopilot supposed to do without Autosteer enabled?…

Unless you mean AEB and side collision avoidance, which use the same sensors, but are *not* classified under “Autopilot” to the best of my knowledge.

Also, isn’t the 40% lower accident rate Tesla claims actually with Autopilot engaged versus without?… I don’t think it would make much sense to make claims about accident rate with the hardware merely installed…

What is the age distribution of Tesla drivers versus overall drivers? How do you think the overall accident rate would be affected if you corrected for that difference? I have no idea if it would be 40% lower, but I will happily bet it’s significantly lower – and the difference may well be more than 40%.

Nor is age the only difference between the Tesla-driver population and the overall population. Teslas are expensive, and that means that buyers are by and large pretty successful people. As such, they tend to have less health problems, less drug addiction, less stress in their lives, perhaps even better eyesight, for all I know.

To get a good handle on this, we should look at Tesla accident rates, ideally comparing cars with AP engaged to cars with AP disengaded, driving on the same road at nearly the same time. Only Tesla has that data, but if it handed it over to a University, was analyzed and published in a peer-reviewed journal, we could get some really interesting insights.

This is not the correct comparison.

If you want to look at the effect of Autopilot on the fatality rate, you need to compare apples to apples: the fatality rate of Teslas driven with Autopilot to the fatality rate of Teslas driven without Autopilot. You can’t compare Teslas with Autopilot to your average beater.

The last time I crunched the numbers, Autopilot increases the per-mile fatality rate by roughy ten times. The safest Tesla is one without Autopilot enabled.

Yeah, never mind finding out the “why” behind the fatalities. Let’s see, the first guy who died was watching Harry Potter while traveling on Autopilot on a back road at 85 mph. The latest fatality was the Model X into the highway split jersey wall that had no marked fog/gore lines at the transition- and he had almost done the same in the same location multiple times prior to the accident. What hard numbers do you have regarding which fatalities were involved with negligence and which were victims of the pure, murderous machinations of Elon’s greed? The devil’s in the details, and we know how little you care about those. Personally, I find no problems driving Autopilot, because I understand the tool and how it works. As I’m sure you do, too, because at last count, don’t you own like multiple Model X’s, a couple Model 3’s, maybe a 1st Gen Roadster thrown in for good measure? You see, you allege to come here to intentionally spread negativity to all those without Tesla experience who mistakenly support the company in order to show them through your vast Tesla experience how they’re ignorantly wrong. But what say you to those like… Read more »

> he first guy who died was watching Harry Potter while traveling on Autopilot on a back road at 85 mph

Ok, he was an idiot. But if AP kills idiots, that is a valid point against it. Of course it must still be weighed against any benefits, but those are also unproven as of now.

So then you’re saying we should tailor our society based on the exception and not the norm? We should punish the masses for the few who don’t want to heed the directions? Sorry, but “saving” idiots from themselves by punishing the responsible is where you and I will have to agree to disagree. If folks wanna disregard the rules, simply disable their cars, but leave mine alone.

The last time I did the math, you killed yourself in your fake Model X when your ran head on with your fake Model 3.

Here, let me show my math:
1 shovel bullpucky + 2nd shovel of bullpucky = 1 pile of bullcrp.

Eleventy Pretend Electrics continued his FUDster campaign:

“The last time I crunched the numbers, Autopilot increases the per-mile fatality rate by roughy ten times.”

It’s literally breathtaking that you actually seem to believe anyone will swallow such obvious bull pucky. It’s also highly insulting that you seem to think anyone reading your comment is clueless enough to believe it.

Crunched what numbers? Show us, please, so we can check your math.

According to Tesla, there are about four times less fatalities per miles driven in Tesla vehicles than the US average. (Some independent calculations based on media reports seem to more or less confirm this claim.) That’s mostly due to their extremely good passive safety, though…

However, Tesla *does* also claim there are 40% fewer accidents on average when Autopilot is engaged.

> Can’t conclude that that is entirely from autopilot, or even that autopilot contributes to it – too many factors involved. It is only too obvious that it isn’t as simple as comparing fatalities per autopilot-driven mile to the national average. There’s loads of systemic differences other than AutoPilot at play here, many of which are known reasons to expect a lower fatality rate in the Teslas. To name a few: – new cars are much safer than old cars – older drivers are much safer than younger drivers – AP cannot drive on all roads or in all conditions; some roads and conditions are much more dangerous than others I believe the latter point is very important. There are far fewer fatalities per mile on highways with physical separation between the directions of travel than on roads lacking this feature. If Teslas are significantly safer than the average new car in a crash, and new cars are much safer than old ones, this fact alone could well be compatible with a much lower fatality rate even if AP contributed *negatively* to safety. Not saying it does, of course – but the data gives very little clue in this regard.… Read more »

Btw I hasten to add that I haven’t seen any better evidence the other way – that AP does not improve safety. All I’m saying is pending a proper, scientific analysis of the data, we have no reason to think we know the answer.

I personally have only seen a single example of an active intervention from AP that seems likely to have avoided a likely serious (but not likely fatal) accident (the one where the Tesla sees the car ahead of the car ahead suddenly braking). Two of the fatalities seem to me rather clear examples of drivers who have been lulled into a false sense of security, despite all the disclaimers and warnings from Tesla and the “beta” status of the system.

To my mind, it seems entirely plausible that AP might, in its current state, do more harm than good. But it also seems plausible to think it might do more good than harm – at least to the large majority who are smart and disciplined enough not to stop driving themselves.

Autopilot tends to be used on safe roads where there are few traffic lights or pedestrians to create dangerous scenarios.

The reason there haven’t been pedestrians deaths due to Tesla on autopilot is not because autopilot avoids pedestrians… It doesn’t!

The crazy thing is that the flood gate of Autopilot cars is just now opening with the Model 3. Given growth rates of more new autopilot equipped cars, on top of the growth in the prior number of Model S/X vehicles. we could see another order of magnitude in autopilot miles in just a few more years.

Other car makers relying on just test fleets can’t ramp their test fleets like this.

But don’t the data sets from each successive million miles of driving have successively less value than the previous data sets, since they’re mostly just going back over the same roads again and again?

As my college American History teacher put it: “Do I deserve regular raises based on the number of years I’ve been teaching? Do I have eight years of teaching experience, or do I merely have one year of experience repeated seven times?” (His emphasis, not mine!)

The amount of data collected by Tesla from its semi-self-driving cars is impressive. But even more impressive, at least to me, is the more advanced self-driving capability shown by Waymo’s fleet of partially or fully self-driving cars.

That is a good point.

To just make up some numbers for example purpose only — From the AI perspective, 1 person overriding autopilot and braking in a corner is insignificant. 100 people do it and the AI sees this as significant and starts to “learn” from it. But 200,000 of the same event wouldn’t teach the AI anything more than 100,000 taught it.

The real benefit in the numbers is in the wide variety of driving patterns that so many drivers in so many cars in so many locations can provide that test fleets just can’t duplicate.

This. Autonomous driving is not about collecting stale miles. It is about reacting to new, unexpected conditions. The past is the past.

Seven fake electrics you are about as dense as a neutron star. What do you think they collect the data for? Just to put on a chart? Maybe they use it to constantly improve their ap systems so they can react to the million of variables that us humans run into every day.

I do understand how you don’t comprehend this since your fictitios model X and model 3 are only ones transmitting made up “stale” data.

Recoil- what you say regarding data collection would get in the way of Seven Electrics negative Tesla spin narrative. But I’m with you, what person poses as a Tesla owner and at the same time runs the company down more than anyone here? It honestly doesn’t make sense, claiming to own multiple, horrible products but espousing to the rest of the world why they shouldn’t own the same product(s)? Dude’s mental.

How much of the data collected is interesting, though? 99% or more is “nothing happens”, then when something interesting happens, was enough sensor data collected? There needs to be a lot of raw data available to improve the classification, for instance.

Yup. My comment was intended to stimulate people to really think about the issue. But perhaps I should have given it a “Devil’s Advocate” caveat, since my comment has a fairly obvious fallacy: Those successive sets of data are going to contain data on at least some new events. They won’t be 100% repetitive.

Yes, the value of the data is logarithmic. Going from say a hundred times more data than others to a thousand times more data, doesn’t mean a ten-fold further improvement, but rather a 50% further improvement. But it’s still a major extra advantage…

Or, there’s always that, too.