How Safe Is Tesla Autopilot?

We don't know if we have accurate fatality numbers for Teslas though. IIRC, the first fatality was unreported for a couple months.

Tesla didn't get telemetry data from the vehicle, so they have no idea what happened yet. (Does Tesla get telemetry data from cars in China at all? Does that get through the Great Firewall? Do Tesla cars in China report to a server in China or back to the US, or what?)

You can see what Mobileye does. It's available as an aftermarket car accessory, and there are lots of videos.[1] There's a technical paper.[2] It's supposed to be able to detect pedestrians. It should be able to detect a stalled car.

But Mobileye doesn't know the steering plan. It warns of possible collisions, but can't warn of every parked car because of its genesis as a driver-warning system. So it may be set up to not warn of collisions which could be easily avoided by steering. So the steering control system isn't told of the avoidable obstacle.

Big arguments over this are now becoming public.[3] The Mobileye CEO says he personally warned Musk not to allow hands-off driving. A former Tesla engineer considered the system unsafe. CNN: "Musk’s philosophy seems to be, 'To make an omelet, you have to break a few eggs.' He is aware that there will be deaths and injuries while Autopilot is in use and is prepared to accept a few negative events in order to prevent many more."

[1] https://www.youtube.com/watch?v=HXpiyLUEOOY [2] http://www.movon.co.kr/download/board.asp?board=blog&uid=865 [3] http://gas2.org/2016/09/17/spat-mobileye-tesla-goes-public/

The "Catch-22" doesn't make a lot of sense, really.

In either case you are driving. Until we legally recognise autopilots as being fully in control of the car, the driver is always driving.

Her statement does make me wonder how relaxing it would really be to drive a car with autopilot, though. If I can't sit back and trust it 100%, I think I'd rather just be fully engaged in driving the car.

Driving on a divided highway is safer then an undivided one. Driving in conditions where you can use cruise control is safer then in ones where you can't. (Snow, ice, sleet, heavy rain, poor visibility.)

I'm not sure where driving on a highway falls compared to city driving.

What I was trying to get at is that the benchmark that Tesla is using as comparison (1 fatality/94 million highway miles) takes into account all highway vehicles. Tesla is a high-end, expensive, luxury vehicle, and the autopilot also does not function in certain conditions (like heavy rain). It would seem to me that if you just took into account other luxury vehicles, and excluded really adverse driving conditions, that the 94 million mile number would probably be higher, and I think that would be a fairer comparison for Tesla to use if they want to claim the car is faster than human driving.

Are you trying to say that you think taking these factors into account would actually make comparables comparitively less safe?

> autopilot: a device that steers a ship, aircraft, or spacecraft in place of a person.

http://www.merriam-webster.com/dictionary/autopilot

I am surprised that Merriam Webster's definition only mentions steering and not acceleration. Telsa's autopilot does have a steering component, so it doesn't seem to be that much of a stretch.

I don't think the average consumer will think about boat 'autopilot' when he hears the term. For most people autopilots are associated with airplanes, because that's where there are actual pilots, not drivers or helmsmen.

?

This name has certainly been HIGHLY vetted by both marketing and legal. I would wager that the name "Autopilot" was created by marketing, not engineering.

> You set bearing, altitude, and speed and the plane will adjust to and keep that course—that's it.

And that's a very useful and safe way to get a plane from A to B, given the amount of route planning and air traffic control in modern aviation. An autopilot can generally provide ample warning of situations it can't handle, and a plane at cruise altitude has quite a bit of room for recoverable error.

Until we have traffic control on our highways that makes sure our cars know about obstacles early enough, the same basic features are not enough to provide the same experience for a car.

The problem with the naming has nothing to do with the technical capabilities, but an understanding of how that will impact human-machine interaction. You are absolutely right that aircraft autopilot is a lot "dumber" in a sense. But there is no ambiguity in the fact that when the pilot turns on the autopilot, he can -- and does expect -- the computer to take over all flight operations.

By naming the system autopilot, Tesla has also created that same expectation with their customers. Not only that, but it has at best turned a blind eye to reports and videos and stories of customers who willfully take their hands of the wheel/sit in the back seat/etc. and at worst encouraged an erroneous public perception of how the autopilot system should be used.

The issues you cited in handoffs with airplanes are far more complicated when it comes to cars, where decisions need necessarily to be taken in a much shorter period of time. But all of that seems really downplayed by Tesla marketing, which instead chooses to aggressively market just how close the system is to full driving autonomy, instead of more like a gentle aid to monotonous highway driving.

The disconnect between their official documentation and their marketing is immense. That is what is potentially dangerous.

There are many degrees of autopilot. It can be something as simple as a connection between a gyrocompass and the rudder to hold heading, all the way up to a fully automatic takeoff, cruise, and landing system.

The issue supposedly is that people have no idea about this, though, and think (because of movies?) that "autopilot" means it does everything. I'm still not sure if I believe it, but if so it has nothing to do with real autopilots.

The difference is that the aircraft autopilot system does not have to be that intelligent. My imagination is that if I engage an aircraft autopilot system I'm free to do other stuff for >= 1minute, because it's enough to simply hold the direction and height. There are no sudden events.

In a car however with the current state of automotive assisantance systems the timeframe for "doing other stuff" is less than 1 second because we have a fast varying environment there, including other cars (or pedestrians, bicycles, ...) taking sudden movements. Even many 5s situations can not be handled by todays assistance systems, like some surface conditions, traffic lights, track changes due to constructions, etc. Which in the end means you have to be constantly be in control of car, which is just not what "automatic" suggests.

You're missing the fact that most consumers are likely totally ignorant of this.

And yet, we have Elon Musk saying that half a million people would have been saved by autopilot, if it were rolled out universally. [1] This doesn't sound like he's pitching a 'maintain speed and heading' device.

So, can we rely on it to make the right decision in a safety-critical situation, or not? (The answer to this question seems to be 'yes, until there's an accident.')

[1] https://electrek.co/2016/07/05/elon-musk-tesla-autopilot-sav...

On top of other comments, don't forget that pilots are highly trained professionals, with a good knowledge of their vehicle and its operation. Very few drivers fall into that category.

On second thought maybe "2nd grader has the wheel" but that's not as catchy.

Hard to find current statistics. This [1] seems pretty believable. Not sure if it's "good" or not, but there you go. It's interesting that, just eyeballing it, mass seems to be the dominating factor. The crown vic is discontinued, but was fairly inexpensive, looks like it was originally $25-30,000. Compared to a C Class, $30-35,000, which had more fatalities.

Of course, this is all 5 years old, so maybe there are better safety features available now.

There are a couple of things that may or may not be believable. Maybe it's possible for one or a few drivers to greatly improve the global traffic system [2]. I believe it, but i can't find a real study that proves it. I do think a handful of self driving cars can smooth out those compression waves, just by driving sanely.

Finally, purely anecdotally, a friend of mine is a truck driver. We get to catch up over dinner every few months. He sees people doing crazy things multiple times a day. He sees the helicopter ambulances on a weekly basis, passing by wrecks in either direction. His assertion, if people would just stay calm and consistent, a lot fewer people would die. It's either someone not paying attention, or being super aggressive.

I, also, would like to see a study. I would bet quite a bit of money, that a highway with only autopilot drivers would have zero wrecks.

I guess the real question is, does sensor failure (device failure, snow, things like that) dominate wrecks? or is it interaction with unpredictable (aggressive, inattentive) drivers?

[1] (pdf) http://www.cbsnews.com/htdocs/pdf/Driver_DeathRates_sr5001_e...

[2] http://trafficwaves.org/