Hacker News Reader

Comment by abalone | Hacker News Reader

I keep wondering what the customer benefit of Level 2 autopilot is if not to lower your attention and relax your mind. Tesla's "out" is that drivers are supposed to retain full attention and oversight of the autopilot system -- but then if you strictly follow this rule, what is the benefit of autopilot?

I can see the benefit to Tesla and future Tesla customers of essentially crowdsourced fleet learning. But what is the benefit right now if you strictly follow the rule of remaining alert enough to intervene at a second's notice?

In previous threads the best explanation I got from Tesla owners was that it frees you from the "details of physically driving" so you "can now supervise instead."[1] That just seems suspect. Supervising autopilot to the degree required to correct sudden mistakes seems, to me, to be probably very close to the mental load of the details of steering yourself.

[1] https://news.ycombinator.com/item?id=17151116

View on HN

but then if you strictly follow this rule, what is the benefit of autopilot?

Indeed. When reading Tesla's statement,

"The Palo Alto-based automaker, led by Elon Musk, has said it repeatedly warns drivers to stay alert, keep their hands on the wheel and maintain control of their vehicle at all times while using the Autopilot system."

...one can't help but think, "then WTF is it good for, if I still have to drive the car myself while using it?"

Supervising autopilot to the degree required to correct sudden mistakes seems, to me, to be probably very close to the mental load of the details of steering yourself.

I'd say the mental load is higher than manually driving --- not unlike being a driving instructor to a student driver who is OK most of the time but occasionally makes a possibly fatal mistake that you need to watch out for and quickly correct. Normally, a car will go in a straight line with no steering input, but the autopilot almost "has a mind of its own" and can steer itself in the wrong direction.

This is subtly different from cruise control, which lets you concentrate on steering and rest your foot on the brake so you can use it quickly if needed. Cruise control doesn't encourage lowering attention, but enhances it.

The NHTSA found a 40% crash rate reduction from Tesla's autopilot. I'd say thats what its good for.

This stat is widely misunderstood. Tesla rolled out Collision Avoidance Assist and Autopilot around the same time. These variables were not isolated by the NHTSA.[1] It is likely that collision avoidance was the major factor in the crash rate reduction. It is even possible that autopilot increased the accident rate.

No doubt collision avoidance is a fantastic feature though.

Addendum: Looks likes the NHTSA has disavowed Tesla’s interpretation of that study. They did not evaluate whether autosteer (autopilot) was engaged.[1]

[1] https://www.cnbc.com/2018/05/02/effectiveness-of-tesla-autop...

> Looks likes the NHTSA has disavowed Tesla’s interpretation of that study...

Wow. This is a story in its own, because that statistics has been touted a million time by the tesla/elon apologists..

so while its possible, Tesla is making changes that bring accident rates down. Everyone here is acting like its certain with absolutely no scientific evidence. Clearly by significantly more than they go up. Its stupid to come down on them like they're murdering people. That IS how you stop all progress.

Furthermore, no one is forcing anyone to buy Teslas. People should be free to do as they choose. I have no doubt, for example, that a corvette, lambo, m3, or convertible all increase certain risks for their drivers by far more than autopilot - which probably decreases their risks.

>This is subtly different from cruise control, which lets you concentrate on steering and rest your foot on the brake so you can use it quickly if needed. Cruise control doesn't encourage lowering attention, but enhances it.

What has lead you to this conclusion? Autopilot is really just advanced cruise control. You can treat it the exact same way by leaving your foot on the brake and leaving a hand on the wheel. If people aren't using it that way, that is certainly something Tesla should try to fix. But I am just not sure I buy the argument that cruise control is inherently a safer technology than cruise control plus auto steering.

Because the fully-assisted driving ("autopilot") removes the need for interaction completely. Your brain will stop spending power on backseat driving, because the brain is really lazy--and clever about being lazy too.

>Because the fully-assisted driving ("autopilot") removes the need for interaction completely.

But it isn't fully assisted and doesn't completely remove the need for interaction. It just is another step forward from cruise control that is more assisted and requires less interaction. I am not sure why everyone assumes the sweet spot for driver assistance ends immediately after adaptive cruise control and before cruise control plus lane keeping.

>Your brain will stop spending power on backseat driving

That is an interesting phrase to use considering backseat drivers have zero responsibility or control of a car but still manage to pay attention.

Backseat drivers pay less attention than frontseat drivers, obviously.

Your characterization of autopilot as “just cruise control” is fundamentally incorrect. Mere cruise control still demands significant and continuous attention and manual control on your behalf (steering). Autosteer makes the leap to fully automatic, which opens the door to not paying attention for long stretches of time.

I am honestly not sure how to discuss this topic when in the same thread Tesla gets ridiculed for blurring the lines between Autopilot and full self driving while people who are critical of the system say things like Autopilot is "fully automatic", "fully-assisted driving", or "removes the need for interaction completely".

From a practical perspective, Autopilot still needs to be monitored. But more relevant to this discussion, the user still needs to interact with the car to keep Autopilot engaged. It is possible to do that while paying very little attention. It is also possible to drive a car with cruise control while paying very little attention. Both systems can be abused. I have already said Tesla should do a better job of trying to crack down on the abuse of Autopilot. But I don't know why it is controversial to say that if used properly both systems can lead to greater safety by allowing the driver to focus on other aspects of driving.

You're mischaracterizing the issue. Autopilot should be monitored but it can go for long stretches without manual intervention (as if fully automatic). Mere cruise control, on the other hand, requires manual steering.

You're just trying to make them sound similar, e.g. "if used properly" or "both can be abused". But there is an ocean of difference. It is vastly easier to ignore a system that is capable of driverless operation for long periods of time.

The biggest problem with this is that Tesla calls it autopilot. Had they called it advanced cruise control or something like that I think people would have used it with more caution. The more accidents that are happening makes me realize how far is Waymo/Google is actually ahead of the rest.

> If people aren't using it that way, that is certainly something Tesla should try to fix.

A good start would be to not call it auto pilot, and rename it to cruise control plus or whatever.

> ...one can't help but think, "then WTF is it good for, if I still have to drive the car myself while using it?"

And why are Tesla burning karma by promoting this technology?

For SpaceX, self-landing rockets are a direct method of reducing costs.

For Tesla, sort-of-self-driving cars are a distraction from the core product and are tainting the brand. Stick to selling top-tier electric vehicles with human-driving, people will still buy them on merit.

To me it's am example of what happens when software developers run the asylum. You can do nearly anything in software, but should you?

>To me it's am example of what happens when software developers run the asylum.

Except if the rumors about turnover in the Tesla software team are true, the exact opposite is the case. The hot goss is that they're under pressure to meet unrealistic goals, not vomiting their own hubris into the product.

In this case, read "software developers" as "Elon Musk," and the metaphor makes sense.

Yes. There is zero chance I'd use a system like this. Regular freeway driving is already boring enough that I worry about losing focus. (I supplement with things like caffeine and audiobooks.)

There's some threshold of interventions/minute that I think humans just aren't very good at. Which is why there are systems like Threat Image Projection, which turns baggage scanning into what is effectively a video game: https://www.rapiscansystems.com/en/products/rapiscan-threat-...

Try it before making such sure if yourself statements.

In my experience it’s LESS boring to drive with autopilot and I pay MORE attention to the actually important, potentially dangerous, things like what’s coming up (you’re actively aware of longer stretch forward, which also means you know you’re coming up to a stretch that autopilot won’t be able to handle well in advance and have plenty of time to disengage) and the cars around it (not watching lanes closely frees you to give all your attention to cars around you).

The actual benefit is to have the car handles itself in situation where you would be distracted regardless.

I engage similar system on my car before fiddling with the attention seeking touch screen, or when I need to give a quick look at the kids behind. Another is traffic jam: a lot of nothing going followed by a lot of things going on - focus on the wrong thing and you can rear end somebody. Minor benefit are some fuel saving if you are a "dynamic" driver, and it can make you more relax in stressful situation like the traffic jam above.

That's just to answer your question. I completely agree with you that the feature is overhyped, not worth paying much for it and actively promoting dangerous driving behaviour. However, it nowadays comes nearly for free as a consequence of other very useful systems: the whole set of emergency braking scenario (emergency, cruise control, safety distance, pedestrian, ...), line keeping assist, blind spot monitoring. It has its uses, if you remain aware of its shortcomings.

But don't Tesla's have Collision Avoidance Assist on by default? That's a distinct feature from autopilot. That's primarily what helps out in situations that you've noted.

Autopilot basically adds autosteering to that. Seems to me that could lull you into more distracted situations that collision avoidance is unable to cope with.

Note: autopilot and collision avoidance both came out around the same time, so it's far from clear that autopilot has resulted in a reduction in accident rates, or if it was mostly just the collision avoidance.

before fiddling with the attention seeking touch screen

...in other words, the feature partially works around some of the problems they introduced themselves by using an overly complex touchscreen UI instead of the traditional no-eyes-needed buttons and knobs.

I haven’t seen a non-touchscreen media player in a new mid- to upper-range car in a long time...

The freeing of the physical part is really the benefit for me. My knees are bad, and sitting in stop and go traffic, or even long traffic lights can actually be painful. Its not bad, but it gets really annoying. Being able to physically move my leg is amazing. You probably do get a lot of that with just adaptive cruise control these days as well, but no real experience with that. The mental load is still identical, but the physical part is way more relaxed.

Exactly: you get that benefit with adaptive cruise control. What autopilot adds is autosteer. What’s the benefit there that doesn’t violate TOS? You are supposed to keep your hands on the wheel and pay full attention just the same.

It moves your mind from micro to macro. Instead of focusing on balancing the car within the lane I can look 500m in front and plan further ahead. I can also spend 2 seconds with eyes off road adjusting the navigation compared to without where you can only spend maybe max 0.5sec. I also feel much fresher after a long trip, constant minor adjustments sounds minor but once you've driven with a good lane keeping aid(not the ping-pong versions) it feels so much more comfortable when the steering wheel is magnetically trying to snap to the center of the lane. It still feels like you are driving but the path of least resistance is always pulling you towards the center of the lane. This would still be considered level 2 self driving, the danger comes when you go to level 3 and really take your mind off the road, those I think should never be allowed on the road.

It still feels like you are driving but the path of least resistance is always pulling you towards the center of the lane.

...except when it doesn't, and you're forced to suddenly comprehend the unusual situation and react before its too late:

https://www.youtube.com/watch?v=6QCF8tVqM3I

That was not 2s but 4s and you could predict that happening even before that. Does not seem like especially dangerous situation when you approach autopilot as GP does.

Totally disagree. Someone was literally killed there. Realistically “2 seconds” could be understating the length of the distraction, plus it takes you a certain amount of time to mentally re-engage and understand what is going on. Especially in a weird situation like that.

A lot can happen in 2 seconds. I hear you and appreciate the detailed description, the best one I’ve read yet. But within it are hints that you are trusting the system more than the terms of service permit.

You feel comfortable looking away a full 400% more than you would under manual control. There is not supposed to be any delta. And for many people, it’s probably more than 2 seconds. Many of these accidents including the recent fatal one happened inside of a few seconds of distraction.

> Supervising autopilot to the degree required to correct sudden mistakes seems, to me, to be probably very close to the mental load of the details of steering yourself.

intuitively, the load seems higher. this is because your brain must now include predictions regarding an additional third party who is not there in cars without autopilot or with autopilot off. it’s similar to how i am more stressed and tired after riding in the passenger seat with a less experienced driver at the wheel than i am at the wheel myself.

I have not enabled it for this reason. I suspect that Tesla might too otherwise they would probably just let people have a 1 month free trial of autopilot so people would pony up the cash.

Compare the number of autopilot crashes, to the number of crashes from cars without autopilot, driven by idiots who are busy staring at the phone screen while driving, you'll realize that autopilot type systems are probably a zillion times safer.

Of course, there's zero data about how many accidents happen because drivers were staring at their phones, so you rarely see articles screaming "<non electric car> driven by an idiot staring at phone screen crashes <into what ever catches attention>".

I was skeptical about Autopilot from day one, but one test drive convinced me that this may not lead to zero accidents, but it certainly is way way better than vehicles that crash when driven by distracted drivers

The benefit for Tesla would be selling more cars because people think they are getting a technology that isn't really ready but Tesla's claiming is (but also claiming not really, read the fine print).

One doesn't have to be quite that cynical. Tesla's goal could genuinely be getting to Level 4/5 autonomy ASAP. My question is whether there is actually any immediate customer benefit to Level 2 autonomy that doesn't technically violate the terms of service.

The benefit is less strain on your feet/ankles/wrists/arms/shoulders etc.

Living in a high traffic area I would appreciate the time I don't have to pump my damn foot up and down every 10 seconds for hours straight while waiting in traffic.

> But what is the benefit right now if you strictly follow the rule of remaining alert enough to intervene at a second's notice?

You have an extra failsafe for those moments where you fail to remain alert. People keep treating these incidents in isolation and failing to account for the fact that human drivers make exactly this kind of fuckup every hour of every day, and lives are lost. Every day.

You get that right? Cars. Kill. People. Every day.

Tesla is shipping systems in the absolute infancy of this technology that are already statistically indistinguishable from human drivers as far as safety goes. And the luddites here want to throw all that in the garbage because of "what is the benefit right now".

Tesla is shipping systems in the absolute infancy of this technology that are already statistically indistinguishable from human drivers as far as safety goes.

That’s a hell of a claim, I’d love to see some serious citations on it.

There is no conclusive study either way. The only real data we have is the two confirmed deaths in Autopilot accidents in over a billion miles driven with Autopilot. That is not enough data to draw a definitive conclusion, but it at the very least shows that there is not a huge problem that makes it dramatically more dangerous.

For comparison the overall casualty rate for the average car in the US is somewhere near 12 deaths per 1 billion miles driven.

I wouldn't say that Tesla's have proven safer, but I think saying they are "statistically indistinguishable" from human drivers is relatively fair at this point.

> over a billion miles driven with Autopilot

No, it's over a billion miles driven by Tesla cars that have the sensor hardware that Autopilot uses. The number of miles driven with Autopilot actually engaged is about 300 million, at least according to the data Tesla has released.

https://electrek.co/2016/11/13/tesla-autopilot-billion-miles...

Look at the date on that article. It is from 2016. Tesla has sold a lot of cars and those cars have driven a lot of miles in the last year and a half. Tesla is certainly over a billion miles in Autopilot driving by now.

EDIT: Here is a Tweet from April 2016 saying 47m miles. [1] Your linked article is from November 2016 and shows 300m miles. That is 250m miles in 7 months. Projecting that forward with zero growth, Tesla would have accumulated roughly 950m miles in Autopilot usage. That isn't including the over 100,000 cars Tesla has sold since November of 2016 or the miles those cars have accumulated in that time. I stand by my statement that Teslas have recorded over a billion miles while using Autopilot.

[1] - https://twitter.com/Tesla/status/718845153318834176

There are other factors that influence fatalities in automobile accidents. You can't legitimately use deaths per miles driven as a proxy for how good a driver Tesla's autopilot feature is.

Sure, which is why I said we had no conclusive data. However, at the very least no red flags have popped up in the statistics. Think of it like flipping a coin 3 times and we got heads all three times. That is not nearly enough to judge whether the coin is fair, but we are probably getting close to concluding that the coin isn't weighted heavily in favor of tails. The sample size is still small but nothing in the statistics is screaming out that Autopilot is a big danger when compared with driving manually.

But we don't need statistics to know that Autopilot isn't as good a driver as a human. By Tesla's own statements Autopilot can't autonomously operate a car, it requires a person to be at the wheel and engaged with driving. By definition this means that autopilot is a worse driver than a human driver as it cannot drive at all.

Well if AutoPilot is statistically saver then a human pilot (I don't know if it is). Then that's a reason to use it.

Why does this get downvoted?

Reply on news.ycombinator.com