Sleeping With Tesla Autopilot May Explain Some Of Tesla's Safety Numbers

Brad Templeton

Recently, there have been a variety of debates and reports of people in Teslas on highways, asleep at the wheel, being driven by autopilot. There have been various videos posted online shot by people in other cars, including this report of police halting a car with a drunk sleeping driver from last year.

When Tesla Autopilot is engaged, it watches for the driver to apply light steering force on the wheel or do other interactions with the wheel.  This is the only way they have to detect if hands are on the wheel or not. Some drivers learn how to apply a constant light force, others tweak every so often.  This has caused great confusion in reports on Tesla accidents which include phrases like "the driver's hands were not on the wheel."  In reality, for many Tesla drivers with their hands firmly on the wheel will not trigger the torque sensor.

If you don't trigger it for a modest amount of time (as little as about 20-30 seconds, according to recent reports) you get a visual warning on the screen. If you go longer, the warning gets brighter. Eventually, at 45 seconds it issues an audible warning. Finally, after 60 seconds, it will begin to slow the vehicle, eventually coming to a stop in the middle of the lane. (This suggests that some of those who have cruised alongside a "sleeping" Tesla driver long enough to take a video may have been punked.)

That's clearly something you don't want to have happened, but it's also vastly better than what happens if you fall asleep at the wheel of a car without Autopilot, which results in an accident a large fraction of the time. In fact, there is a strong suspicion that a very significant number of car accidents and fatalities are caused by falling asleep at the wheel. Alas, there is no blood test to perform on a crashed driver to find out if they were asleep, the way one can test for drinking.   We do know that 60% of fatal accidents are single-car accidents, where a car simply ran off the road, and only some of them involve alcohol.

You clearly don't want to fall asleep when using a driver assist lane-keeping/ACC system like Autopilot. These systems may drive for an hour without needing any intervention and seem very impressive, but that rate would mean a dangerous high-speed accident one time in 60.  In addition, stopping in the middle of the highway is a poor idea. At the same time, if you do end up drifting off, you are much better off if you have such a system on. (This leads to a paradox I'll discuss below.)

Tesla safety numbers

Tesla releases safety numbers every quarter. Their most recent report reads:

In the 2nd quarter, we registered one accident for every 3.27 million miles driven in which drivers had Autopilot engaged. For those driving without Autopilot but with our active safety features, we registered one accident for every 2.19 million miles driven. For those driving without Autopilot and without our active safety features, we registered one accident for every 1.41 million miles driven. By comparison, NHTSA’s most recent data shows that in the United States there is an automobile crash every 498,000 miles.

At first, the numbers sound very impressive.  However, in spite of repeated attempts, Tesla has declined to explain the difference between an "accident" and a "crash." The NHTSA number involves crashes reported to the police. There is speculation, which Tesla refuses to clarify, that "accident" might mean an airbag-triggering crash. There is strong reason to suspect different things are being measured because even with the ADAS tools in a Tesla, drivers not using Autopilot should not have 3 times as good a record than average.

Another complication is that Autopilot is marketed as only for highways, and while it works on some non-highways, the bulk of its use is there. The accident rate on freeways, per mile, is drastically lower. Some research suggests as much as 2-3 times lower. The severity of highway crashes is greater, but the frequency per mile is much lower. Tesla has declined many requests to disclose how many of the Autopilot and non-Autopilot miles listed above are on freeways, rural highways and urban streets, and until they do, the numbers they publish do not provide useful information.   While they might, as Tesla implies, suggest that driving with Autopilot has an overall superior safety record than driving without it, we don't know enough to say. In addition, the NHTSA statistic is for all roads.

Nevertheless, if the accident rate with Autopilot on the highway is truly better than the rate for non-Autopilot on highway – an "apples to apples" comparison – the sleeping factor may be more than sufficient to explain it. People who fall asleep in regular cars (or Teslas not using Autopilot) are probably getting in a lot more accidents than those using the system.  (It should also be noted that many cars today have collision warning and prevention systems and lane departure warning systems which will also awaken drivers or even prevent an accident. These systems will act when the car veers from the lane or is about to hit a vehicle in front, while the Autopilot does most of the driving task so that those things don't happen in the first place, but both will reduce accidents.)

The paradox

In general, it sounds like it's a good idea if driving while drowsy, or impaired, to have Autopilot on. Yet at the same time, it's a terrible idea to drive while drowsy or impaired. The paradox is that the increased level of safety could lead people to take the risk. If you can't keep your eyes open, you should pull over – yet many people don't. The existence of Autopilot can easily lead to somebody being less likely to pull over, or worse, less likely to pull over when drunk. We all know we aren't supposed to drive in these states, but we do, and it's probably the cause of half of all serious accidents, particularly on weekend evenings. More study is needed to learn how the existence of ADAS and Autopilot changes the risks people take, to figure if it's an overall safety loss or win.

Read/leave comments at this site.

">
Brad Templeton

Recently, there have been a variety of debates and reports of people in Teslas on highways, asleep at the wheel, being driven by autopilot. There have been various videos posted online shot by people in other cars, including this report of police halting a car with a drunk sleeping driver from last year.

When Tesla Autopilot is engaged, it watches for the driver to apply light steering force on the wheel or do other interactions with the wheel.  This is the only way they have to detect if hands are on the wheel or not. Some drivers learn how to apply a constant light force, others tweak every so often.  This has caused great confusion in reports on Tesla accidents which include phrases like "the driver's hands were not on the wheel."  In reality, for many Tesla drivers with their hands firmly on the wheel will not trigger the torque sensor.

If you don't trigger it for a modest amount of time (as little as about 20-30 seconds, according to recent reports) you get a visual warning on the screen. If you go longer, the warning gets brighter. Eventually, at 45 seconds it issues an audible warning. Finally, after 60 seconds, it will begin to slow the vehicle, eventually coming to a stop in the middle of the lane. (This suggests that some of those who have cruised alongside a "sleeping" Tesla driver long enough to take a video may have been punked.)

That's clearly something you don't want to have happened, but it's also vastly better than what happens if you fall asleep at the wheel of a car without Autopilot, which results in an accident a large fraction of the time. In fact, there is a strong suspicion that a very significant number of car accidents and fatalities are caused by falling asleep at the wheel. Alas, there is no blood test to perform on a crashed driver to find out if they were asleep, the way one can test for drinking.   We do know that 60% of fatal accidents are single-car accidents, where a car simply ran off the road, and only some of them involve alcohol.

You clearly don't want to fall asleep when using a driver assist lane-keeping/ACC system like Autopilot. These systems may drive for an hour without needing any intervention and seem very impressive, but that rate would mean a dangerous high-speed accident one time in 60.  In addition, stopping in the middle of the highway is a poor idea. At the same time, if you do end up drifting off, you are much better off if you have such a system on. (This leads to a paradox I'll discuss below.)

Tesla safety numbers

Tesla releases safety numbers every quarter. Their most recent report reads:

In the 2nd quarter, we registered one accident for every 3.27 million miles driven in which drivers had Autopilot engaged. For those driving without Autopilot but with our active safety features, we registered one accident for every 2.19 million miles driven. For those driving without Autopilot and without our active safety features, we registered one accident for every 1.41 million miles driven. By comparison, NHTSA’s most recent data shows that in the United States there is an automobile crash every 498,000 miles.

At first, the numbers sound very impressive.  However, in spite of repeated attempts, Tesla has declined to explain the difference between an "accident" and a "crash." The NHTSA number involves crashes reported to the police. There is speculation, which Tesla refuses to clarify, that "accident" might mean an airbag-triggering crash. There is strong reason to suspect different things are being measured because even with the ADAS tools in a Tesla, drivers not using Autopilot should not have 3 times as good a record than average.

Another complication is that Autopilot is marketed as only for highways, and while it works on some non-highways, the bulk of its use is there. The accident rate on freeways, per mile, is drastically lower. Some research suggests as much as 2-3 times lower. The severity of highway crashes is greater, but the frequency per mile is much lower. Tesla has declined many requests to disclose how many of the Autopilot and non-Autopilot miles listed above are on freeways, rural highways and urban streets, and until they do, the numbers they publish do not provide useful information.   While they might, as Tesla implies, suggest that driving with Autopilot has an overall superior safety record than driving without it, we don't know enough to say. In addition, the NHTSA statistic is for all roads.

Nevertheless, if the accident rate with Autopilot on the highway is truly better than the rate for non-Autopilot on highway – an "apples to apples" comparison – the sleeping factor may be more than sufficient to explain it. People who fall asleep in regular cars (or Teslas not using Autopilot) are probably getting in a lot more accidents than those using the system.  (It should also be noted that many cars today have collision warning and prevention systems and lane departure warning systems which will also awaken drivers or even prevent an accident. These systems will act when the car veers from the lane or is about to hit a vehicle in front, while the Autopilot does most of the driving task so that those things don't happen in the first place, but both will reduce accidents.)

The paradox

In general, it sounds like it's a good idea if driving while drowsy, or impaired, to have Autopilot on. Yet at the same time, it's a terrible idea to drive while drowsy or impaired. The paradox is that the increased level of safety could lead people to take the risk. If you can't keep your eyes open, you should pull over – yet many people don't. The existence of Autopilot can easily lead to somebody being less likely to pull over, or worse, less likely to pull over when drunk. We all know we aren't supposed to drive in these states, but we do, and it's probably the cause of half of all serious accidents, particularly on weekend evenings. More study is needed to learn how the existence of ADAS and Autopilot changes the risks people take, to figure if it's an overall safety loss or win.

Read/leave comments at this site.

I founded ClariNet, the world's first internet based business, am Chairman Emeritus of the Electronic Frontier Foundation, and a director of the Foresight Institute. M...