GM Cruise to delay commercial launch of self-driving cars to beyond 2019 | Hacker News Reader

GM Cruise to delay commercial launch of self-driving cars to beyond 2019

reuters.com

100 pointsdrkimball174 comments

Cruise cars drive like a drunk grandma stopping short all over the place and regularly failing to take a protected turn at a green turn light because some pedestrian has a foot edging into a crosswalk.

The biggest safety issue associated with these cars is that everyone else honks and swerves aggressively to evade them when they stop short for dumb reasons like someone in the road waiting in front of their car for traffic to pass so they can get into their car —- these are actions I witness multiple times every day and if you look at every accident report associated with these cars you’ll see almost every case seems like a driver frustrated about being stuck behind it trying to swerve around it.

What isn’t recorded is all the accidents involving the cruise cars that don’t actually impact the cruise car. Like when someone swerving around a cruise car side swiped my neighbors car on Fulton. I almost saw a biker on 14th street by Valencia get hit in a similar manner when someone trying to evade the cruise car in the left lane swerved quickly into the right lane while the biker was avoiding a car in the right lane trying to turn right onto Valencia.

They almost certainly don’t collect statistics on accidents caused by other drivers who just want to get around these cars that clearly don’t need to be tested in a way that makes everyone’s commute just a little bit more obnoxious.

> The biggest safety issue associated with these cars is that everyone else...

...drives dangerously and therefore illegally?

That's an indictment of human drivers, not self-driving vehicles.

So start aggressively ticketing all of the reckless humans and maybe that behavior will stop.

If you have to swerve, then you're not following at a safe distance. If you swerve without having to, that's also recklessly dangerous. Both behaviors _should_ be penalized.

There's more to it than that. Traffic is as safe as it is (I won't call it flat-out "safe"), in no small part because people can anticipate other drivers. A _predictable_ driver is a safe driver. Having self-driving cars randomly stop short is a quick way to rack up a lot of "not legally at fault", but perhaps morally at-fault, accidents.

FWIW, if you've ever done car/bike track days, this is a huge reason why it's so safe. People are split into groups according to skill level, with restrictions on passing (normally "only on the straights") for the lower groups. Much of the "skill level" here is actually just being predictable, not being outright fast. Someone can be in a Miata getting passed with a 50mph speed differential on a straight _because everyone is being predictable_. That same Miata might pass the same car with a huge speed differential at 90 mph in a turn, and it works because everyone is predictable.

>Having self-driving cars randomly stop short is a quick way to rack up a lot of "not legally at fault", but perhaps morally at-fault, accidents.

And I'm pretty sure, in most jurisdictions, slamming on the breaks unexpectedly and for no good reason, will get you at least a portion of the liability for an accident. The person who rear-ends you will get some too, but its not always true that you can't be at fault when you get rear-ended.

As you imply, no safety system should rely on others doing their job perfectly. That is an inherently unsafe system.

If there was an accident and the cops ask you "why did you break suddenly?" and you say something reasonable that's not "to commit insurance fraud" you probably won't be at fault. It's the responsibility of the person behind you to keep a safe following distance.

I have had the "pleasure" of driving alongside and behind Cruise's vehicles on many occasions here in San Francisco. They will stop dramatically in the middle of a block (empty, no traffic, no pedestrians), edge forward, dramatically stop again, go halfway through the intersection and again...dramatically stop.

It feels to me like whoever has written their braking algorithm is using bang bang control.

Sure, but that an issue of evidence. Google, Uber, Tesla, etc. have source code, records, emails, reports, etc. that can potentially prove their cars just break because they aren't good enough drivers. And they have deep enough pockets that lawyers will go after them.

1000% sure Google will end up writing a check if someone dies because they rear-ended a google car that suddenly hit the breaks.

Your proposed solution is obviously untenable. The danger I’m describing is that not all cars are self-driving and the way human drivers react to these cars dramatically increases the danger around these cars. Drivers in SF may already know what to expect, but in every new city they roll out in, the human drivers will react rashly to these cars as they are frustrated with the unpredictable, inhuman, and incommunicable (no eye contact or hand waving) decisions they make.

> human drivers...dramatically increases the danger

I know you're putting a gradual transition into context, but I see you say this ^, and my reaction is to suggest that we take human drivers off the road who recklessly endanger others instead of giving them a pass with the vehicular equivalent of "boys will be boys".

I am so down for a car-less city

Having traffic be predictable is far more important than the specific rules it follows. Human driven traffic follows rules, including a massive context specific set of rules about when it is reasonable to break the written rules.

Drivers that adhere to the letter of the laws and totally disregarding the unwritten rules make terrible drivers because they do not follow the full set of rules that the majority of human drivers follow. All current autonomous cars drive like this. A small but infuriating subset of human drivers behave like this too. Generally it's expected that you pass these obstructions when possible and safe enough to do so (legal or not, though what's safe is usually legal too).

What people consider safe varies somewhat between people. Anyone will pass a tractor on a country two lane. Only the most bold will pass a Cruise in traffic. Of course there's a danger zone around it.

The current system is designed for humans and if we're ever going to automate it the autonomous cars need to be able to coexist with the human cars without causing problems. Boys will be boys isn't an excuse because it doesn't need to be. One of the engineering requirements is to get along with the boys.

Well I do not think the issue is about following the unwritten rules. The issue is for a company to admit "yeah we're willingly programing/training our vehicles to break the law". Sure, it might be actually less dangerous to do so, but I'm not sure their PR department will be happy about it.

I fully expect you are correct. I'm sure the engineers riding along are well aware how overly conservative they're being.

are the engineers "over-conservative"? or have human drivers become inured to what should have been considered an unacceptable amount of driving risk?

What's "acceptable risk" is basically a matter of social consensus so anything human drivers (as a group) are ok with is de-facto acceptable.

I think the engineers know the software drives un-humanly but they can't change it because driving like a dick within the letter of the law results in far less legal liability than driving like a human and breaking the law.

> Human driven traffic follows rules

This is wishful thinking and clearly false. If it were true, there wouldn't be any at-fault crashes. The only accurate claim you can make is "human traffic sometimes follows rules, and the rules that it sometimes follows change regularly from person to person and moment to moment". That's not a defensible state.

Implicit in your mistaken description is a pretense that humans are psychic and that we can predict what other people will do, but we aren't and, despite believing otherwise, we can't. That's why defensive driving is so important.

>This is wishful thinking and clearly false

No it's not, but you'll likely never see it that way let alone admit you're wrong on the internet.

>If it were true, there wouldn't be any at-fault crashes.

Computers follow rules perfectly 100% of the time yet we still have software bugs.

>The only accurate claim you can make is "human traffic sometimes follows rules, and the rules that it sometimes follows change regularly from person to person and moment to moment".

You understand rules to be flexible and subject to some interpretation (like laws, not like a firewall config). Context is important. The rules that govern human interaction are complex.

>That's not a defensible state

This comment proves otherwise.

>Implicit in your mistaken description is a pretense that humans are psychic and that we can predict what other people will do, but we aren't and, despite believing otherwise, we can't

Humans don't need to be psychics to accurately predict what other drives will do. Humans are really good at pattern identification and picking up on subtle cues.

>That's why defensive driving is so important.

Being able to make accurate predictions of the future state of what is around you based on the current and past state of what is around you is far more helpful than defensiveness for any amount of defensiveness.

The behavior by AI vehicles fielded by Cruise and similar is a glaring example of that while defensiveness is great at avoiding liability it does not necessarily make the road around you safer (maybe even worse, as many comments anecdotally point out). This is the main gripe people have with AI and human driven vehicles that practice defensive driving to the point of absurdity. Defensive driving without human level ability to predict what other traffic will do is not sufficient for the kinds of situations that are regularly encountered in urban areas.

Most people will never get in a car crash in their entire lives (bumping into an object in a parking lot notwithstanding). I think that speaks for itself.

> Computers follow rules perfectly 100% of the time yet we still have software bugs.

At-fault crashes are not analogous to software bugs. The computer didn't look at its phone when your code has bugs. Your off-by-one error didn't happen because the computer changed lanes suddenly without signaling.

Computers also don't follow predictable rules perfectly 100% of the time. That's why we have things like ECC memory and other single event upset detection.

You're working from an ideal model of the way you think things ought to work while ignoring the reality of what actually happens.

I agree with you that people should be ticketed for reckless driving, but I believe you're missing the point. If adding these cars to the road causes more accidents, regardless of who is at fault, they are making the roads more dangerous.

Just like enforcing drug laws doesn't seem to increase public safety (I welcome data one way or the other, but my understanding is it actually is a significant contributor to decreasing public safety), maybe we should be honest with ourselves about human behavior and write our laws to actually cater to that.

If humans reject self-driving cars, making laws protecting them isn't going to make people accept them. Make the self-driving cars actually acceptable to humans and then you don't even have a problem you try to solve with regulation.

Wouldn't it depend on the type of accident? From what I understand the most common case is that an autonomous vehicle will not proceed when a human thought it would (eg. at an intersection), resulting in the human hitting the back of the autonomous vehicle at low speed. That type of accident is preventable through patience and not very severe. It seems fine to me to have more of those if it results in any decrease in more severe accidents.

It was a hypothetical - it has nothing to do with the data because nobody is sharing the data. You can reinterpret it to mean "more dangerous" or however you like. The main point is culpability is irrelevant if the goal is safety.

Plus the SDC has cameras and sensors to capture other drivers being reckless and that data can be used as evidence to prosecute those drivers.

IMO it is less desirable to be a person who incites rage as compared to a person who becomes occasionally enraged.

Driving is strange in that there are upper bounds, but rarely lower bounds. i.e. a minimum speed is a rarity (and even more rarely enforced).

If I made the rules then we would have speed targets (not limits) and people would be penalized for driving outside of a +/- 10% range. After all, it's not speed itself that kills. It's a large and sudden speed delta.

Ideally, the speed targets would also be variable and data driven.

The higher the speed, the higher the speed delta will the road and all stationary objects.

That's true, but I'm not advocating ludicrous speeds. I feel like people rarely crash into stationary objects when they're traveling on a highway and are more likely to rear-end someone entering their lane traveling significantly slower than the flow of traffic (or swerve around them and accidentally hit someone else). Increasing the flow rate should in theory also cut down on the amount of traffic on the road at point in time; making crashing into others less likely by default.

Also, crash your car off the road or into walls all you want. Just don't crash into someone else please.

Stopping suddenly for no good reason is also reckless behavior. Issuing a bunch of citations is not a remotely reasonable solution. These cars are clearly not ready to be tested in a place like San Francisco.

This "humans are the bugs" thing is indicative of the disconnect between reality and the pre-programmed car proponents.

https://news.ycombinator.com/item?id=20462051

One nice feature of these cars is that they have always-on 360-degree video, which should make it trivial to determine fault in the case of accidents. Even if the self-driving car itself wasn't hit, the video could be provided to shed light on what happened around it.

So if you're reckless, swerve around it because you're frustrated, etc. be prepared to be held at-fault and watch your insurance rates skyrocket.

Yes, the data collected in every accident can be used to improve the self driving car too, but probably not increase their predictability. it’s a gray moral zone to suggest that emotional, frustrated people are totally at fault for the statistically increased recklessness around self-driving cars being tested during rush hour. It’s bad enough with the Uber and Lyft drivers who clearly drive like they have no idea where they are or where they are going; but at least do so predictably.

I'm not sure what it implies statistically, but I've now seen the aftermaths of two fairly significant (ie, left in undriveable condition) cruise car collisions just on my block in SF in the past few months. In fact they're the only automobile collisions I've seen on my block in years. No idea of cause or responsibility, of course.

N=2 implies nothing statistically.

So, human drivers are terribly impatient and reckless? This is one of the reasons we need to take cars out of human hands. Cruise cars could integrate better with traffic, no doubt, but you can't blame Cruise for humans being impatient and shitty drivers.

Some are, not all. And even then, they maybe not always are. Please, don’t generalise. We have laws and ways of dealing with reckless drivers.

It's totally irrational but living near the Mountain View location of Waymo makes me hate having the self driving car development process. They add to the traffic in a very congested area (making left turns on Rengstorff I assume to regression test, but this stops traffic for extended periods), then show up in low traffic suburban areas that would otherwise be quiet and clear of traffic ( for some reason they love to use through streets in Los Altos for training and my 300 yard long street at all hours including at night - I can recognize the whine of those cars now), and when behind one, a driver notices how carefully the self driving car has to consider the possibilities and how slowly it acts on decisions - like a 90 year old person reaction time, and giving pedestrians far away from the crosswalks the benefit of the doubt - they could run out into the street and the self driving cars I've been behind seem to assume this is going to happen every time - I can at least look at the person's body language and head and where they are looking. Of course this is only anecdotal from driving errands around Mountain View and cycling around Palo Alto/Los Altos.

it used to be way worse when waymo had they had those tiny electric cars driving similar routes but below 30mph

At least those were adorable, they were like local wildlife.

A very conservative lower bound is that every day we don't have self-driving cars costs the country $1 billion in lost value. Every single day.

> A very conservative lower bound is that every day we don't have self-driving cars costs the country $1 billion in lost value. Every single day.

And every day we don't have solar cells are 2000% more efficient than current technology costs us tens of billions in lost value. What exactly is your point?

The point is that the inconveniences being described by stevenwoo are utterly trivial. It's like listening to someone who lived in Cape Canaveral during the '60s complain that they dislike the whole Apollo program because one of the engineers once left a 5% tip in a local restaurant. Acknowledging that this feeling is irrational doesn't make it a useful HN comment.

I see your point. I'm left wondering if there aren't real impacts to testing this technology in the real world though. Like if there have been any deaths due to emergency responders being stuck behind these vehicles or something similar.

well there is a known upper bound on solar cells efficiency. and even then we are pouring all the moneys into improving solar cell efficiency

It's even higher for the cold fusion we're missing, and teleportation technology.

If you think those three techs have comparable likelihood of existing in our lifetimes, I'd quite seriously like to make a bet with you.

Source?

Your reaction is not irrational at all. This sort of abuse of public space to test technology that is nowhere close to being ready should make you angry. They are trying to progress more rapidly by externalizing the costs of experimentation onto the public. They've hit a plateau and don't have a clue how to get beyond it, but they've got billions of dollars of pressure on the backs of promises they made, so they can't admit their fundamental approach is unsuitable to reach what is actually a questionable end-goal in the first place--namely, self-driving taxis. Why is that the first problem they are trying to tackle? That's the hardest problem to solve. I think it all goes back to Uber struggling to defend its business model. To justify their nine- and ten- figure quarterly losses, they had to promise a world where the labor costs would go away (how driverless Ubers could ever pay back all the money they've lost, I have no idea--if they replace contract labor with their own capital and the maintenance involved... how is this a profitable company?). They convinced their investors it was possible. And then everyone else assumed it was possible and bought into that same fantasy. I hope we're finally waking up. But those billions have left a major hangover and no one is thinking clearly quite yet.

>Cruise cars drive like a drunk grandma...

Don't forget the part where said drunk grandma's proponents say she's a great driver because she has so few at fault accidents. ;)

As a cyclist, I can't wait for self-driving cars.

I'll take a small chance of being accidentally killed as a systemic, correctable mistake over the small chance of being killed by the 5% of people who are consistently impatient, inconsiderate dangerous jerks.

I'm not a cyclist, but I want to be. I don't bike because doing so is way too dangerous with the number of drivers who are paying extremely close attention to their phones and not to driving. And yeah, some drivers are rude too.

I think the most dangerous time to be a cyclist or pedestrian is right now. Smartphones are ubiquitous, distracted-driving enforcement is very weak (non-existent in my area), and self-driving is alpha quality.

You can always mountain bike!

There's some truth to that, particularly the point about phones, but FWIW most cities have drastically increased the number and quality of bike lanes in the last 10 years.

SF is a completely different experience for bikes than it was 5 years ago, and 10 years before that it was even more different.

Ditto for NYC. You can ride up and down Manhattan in big separate big lanes very quickly.

It's obviously a personal choice, but I think cycling is a safe choice now. When I started 10+ years ago I would sometimes get off the road when conditions/traffic looked dangerous, and you should always keep that option in your back pocket. (i.e. don't insist that 100% of the route has to be bikable.)

But I don't feel the need to do that very much now. The exception is when there's unusual traffic and no bike lane.

It's even safer now that there are more cyclists on the road. Drivers do take some time to get used to cyclists.

I think the anecdote applies for some cities committing to Vision Zero plans - for example NYC has data showing fatalities decreasing (page 13: https://www1.nyc.gov/assets/visionzero/downloads/pdf/vision-... ), though injuries might be the same or increasing ( https://nyc.streetsblog.org/2019/02/04/bloody-january-cyclis... )

But it isn't perfect - SF fwiw has had more pedestrian fatalities this year though they've had reductions in the prior couple years ( https://www.visionzerosf.org/wp-content/uploads/2019/07/06.2... )

Also average vehicle miles travelled in the US is up ( https://fred.stlouisfed.org/series/M12MTVUSM227NFWA/ ), and so are nationwide pedestrian fatalities: https://usa.streetsblog.org/2019/01/23/the-bible-belt-should...

Ride-hailing usage over transit may be part of the problem: https://steps.ucdavis.edu/new-research-ride-hailing-impacts-... ) .

I agree that it's dangerous, but I love biking too much not to – usually just as a commuter. I always just assume every driver has either: 1) not seen me and is going to accidentally kill me, or 2) seen me and is going to intentionally try to kill me. Joking aside, if you learn how to assert yourself on the road, take a lane when you need to, and pay attention to what potentially threatening vehicles are up to, I think you can drastically reduce the chances that you are involved in a harmful accident. Always wear a helmet, use really bright lights at night, don't ride on sidewalks, and all that jazz, too.

On my commute, if I try to "take a lane..." I'll end up as bug splatter on the grill of a 2 ton car. The only safe way to ride to work is either using the bikepaths (which don't typically follow roads so take 3x as long, or ride on the sidewalk.

I agree that taking the lane can be difficult and I can only imagine in USA how much harder this is. Here in the UK the number of times I've had people squeeze by me at speed, rather then waiting to overtake safely - but I do find it happened far far less once I got out the gutter and into the road.

I could be wrong, but I believe the statistics show you're much more likely to be killed in that car than on a bike.

From what I've heard about the studies done, you're still liable to live longer with the extra exercise from biking to work even given the increased odds of being killed in traffic.

I think you have way too much faith in self-driving cars. We’re very far off, and anyone telling you otherwise is leaving out important holes.

As a pedestrian, I can't wait for self-driving bicycles ;-)

As a bicycle, I can’t wait for self pedestrianing drivers.

You joke, but a lot of auto-vs-bicycle collisions can be attributed to unexpected actions by the bicyclist, many of which probably couldn't have been compensated for by the automobile driver, perhaps not even by an automated one. Obviously the bicycle always loses in these, but the cause isn't automatically bad behavior by the automobile. (though undoubtedly it often is)

"Systematic, correctable mistake" is not the way to describe errors made by AI algorithms, and definitely not in the context of an engineered system containing many algorithms and sensors.

In general, AI algorithms are famously opaque and hard to debug.

It's too long to explain here, but here are some links:

https://en.wikipedia.org/wiki/Explainable_artificial_intelli...

https://www.wired.com/story/greedy-brittle-opaque-and-shallo...

https://blogs.wsj.com/cio/2018/05/02/companies-grapple-with-...

https://thenextweb.com/artificial-intelligence/2017/04/23/ar...

https://towardsdatascience.com/need-for-explainability-in-ai...

https://www.forbes.com/sites/cognitiveworld/2018/12/20/geoff...

-----

BTW I'm also a cyclist. Human drivers can be annoying, but I find that they do act predictably. I've been cycling 5+ days a week in SF for 10+ years. I find that the drivers are worse when I bike in cities where bikes aren't as prevalent. Humans do learn. When you get into a conflict, there are emotions, and emotions are mechanisms for learning :)

It seems like you think that an AI is "just" a computer program, where the errors ARE indeed systematic and correctable. But AI is fundamentally different.

It's still systemic and correctable. The fact that the ML is somewhat opaque doesn't really matter here. You correct it by generating test cases and adding more training data related to those cases. The issues are still systemic and correctable.

And 80% of people who are texting and distracted while driving.

How long before the self driving and sentient AI cars are complaining to each other about all the stupid human cyclists, their unpredictable behavior, their inefficient mode of transport and poor decision making models?

This is so strange. As a cyclist, I would be very worried about self-driving cars since cyclists - to the eyes of an AI algorithm, don't look the same.

You can have road bikes, mountain bikes with similar sitting positions, or recumbents, which will need a very different training profile, or tandems, etc

Bikes are severely underrepresented in the list of things that car automation companies care about at the moment, so any self-driving car on the road today might misidentify bikes.

Bikes also often ride in positions (from the perspective of the car) that can be unexpected. Some can be as close to the side of the car that it can cover the full field of view of a side-looking camera for example. How do self-driving algorithms work when encountering such cyclists?

I sure don't want to find out myself - I would hope car companies are doing tests on closed courses with cyclist mocks or inputs.

> Bikes are severely underrepresented in the list of things that car automation companies care about at the moment

How can you possibly know that?

They don't even do testing with pedestrians ...

And you know this how?

The same logic applies to human drivers - various bike sizes and types are harder to see and adapt to. If everyone rode a Jump bike, bright red and upright with blinking lights, it'd be easier to see them all quickly. But I'm constantly seeing tiny mopeds, recumbents, kids' bikes -- it's all harder to spot than a Jump bike, even for a human driver.

So, why fall back on the status quo of a human to spot them? If SDCs are proven to do it better, then I will definitely cycle more once there are more SDCs. It's all about evidence, show me the evidence SDC's are better, then I'll be 95% on board.

The likelihood you'll be misclassified as an errant shadow and run over is extremely high for the next 10+ years of self driving machines. On the plus side, the faulty parties will have deep pockets you can help yourself to if you survive.

That's a danger with any vision based system, including people, but much less of a danger if you use lidar like GM Cruise does.

Just because a sensor can see you doesn't mean the decision maker will take the appropriate action.

Oh please. You're propagating these falsehoods about autonomous vehicles that don't require much thinking to disprove.

1. The classification shouldn't happen from a single frame. In order to be misclassified as a shadow, the system would have to get the classification incorrect at various distances and you're probably moving too, so it would need to be across many frames.

2. This suggests that one sensor misclassifying will be sufficient for a deadly situation. But autonomous vehicles all use multiple sensors. They could even theoretically have multiple classification models running. One misclassification will not result in accidents.

3. Humans driving around effectively blindly is already an extremely dangerous situation.

4. This discussion is often limited to places where humans are distracted but overall trying to drive properly. In many parts of Asia (Korea, Taiwan, etc.) drivers very willingly disobey signals and are not at all deterred by the risk of hitting a pedestrian through their own impatience.

Humans are very dangerous behind the wheel. Hypothesizing about corner cases where autonomy might make a mistake is missing the forest for the trees.

Alas, at least where I am cyclist seem to have the least amount of respect for all others on the street and for the traffic regulations.

get a stationary, roads are for cars

The sad thing is that vapourware and the promise of technology to easily solve our problems has real impacts on delaying progress by other means.

Cars are the number one killer of children in America. Political action to reduce this number by redesigning cities to limit human/car interactions and providing safer transportation alternatives is stymied by opponents to who point to assuredly safe autonomous cars as being just around the corner.

See also Hypertube vs high speed rail and carbon capture vs carbon taxes.

To be sure, the car lobby has existed and worked steadily to make cities more car-centric since long, long before the self-driving car people came around...

We (in the US) have failed to make safer cities for about a hundred years already, so this sounds a bit like blaming people trying to build new fences for the existence of the wolves.

Your examples just show our political system is corrupted by big money, it's not about tech and vaporware.

-Cars have power due to historical considerations: car lobbyists and rich car owners getting their environments designed for their own car-use.

-Hyperloop has almost nothing to do with high speed rail failing in the USA, it is a brand new concept and there were decades of destruction of trains by the auto industry and by the US's poor geographic make-up for trains.

-Carbon capture is just a fake-out by fossil fuel industry, there is no economic potential for CCS. Carbon taxes can't even pass in liberal WA in two referanda. Carbon taxes are simply extremely unpalatable, due to massive fossil-fuel lobbying against them and to subsidize fossil fuels. Therefore, most Americans are ignorant of just how much their fossil-fuel habits truly cost, since their other taxes and politicians ensure they are cheap and readily available with no taxation. Big moneyed interests have mortgaged our planet's future to keep gas at $2.50/gallon.

Wise move. Self-driving cars are likely to need a year more of testing for the next quite-a-few years.

Yeah it's been pretty entertaining to watch this shoe drop. I feel like 2 years ago the self-driving car manufacturers were releasing crazy deadlines, like full self-driving in 5 years. But they seemed very confident, so I believed them, despite my skepticism. Now they're walking all that back and removing any definite-ness from their timelines, which falls back in line with my picture of reality.

> I feel like 2 years ago the self-driving car manufacturers were releasing crazy deadlines, like full self-driving in 5 years.

Everything is always five years away. That's the sweet spot for getting lots of funding while allowing time for the wow factor to wear off so you can backtrack on what you really intend to deliver (if anything at all).

This isn’t unique to car AI, every software project looks doable at the start, you can even get “90%” there where it looks and feels like a real solution. But then real life usage testing starts or the implementation process exposes a thousand small problems you didn’t except in the last 10%, or your own original design wasn’t flexible enough for some constraints requiring a significant rethinking of that 90%. Just ask Jeff Hawkins.

This is going to be a huge business opportunity for car manufacturers and software companies, two industries with access to plenty of capital. So the not being left behind risk (FOMO) is naturally going to be high.

Combine that with potentially saving hundreds of thousands of lives from car accidents and being able to look at your phone instead of driving (freeing up countless human hours) and you have the recipe for extreme optimism.

The only important question is whether x company overpaid in the short term for acquisitions or scaling up ... and/or lack a runway to really achieve their goals. Otherwise it’s pretty par for the course, especially for software/innovation at this scale.

It’s quite possible we’ll hit these goals sooner now that industry went full bore on solving these problems and we’ll benefit as a society in many ways besides transportation from the AI work.

Most importantly, who cares if some pension funds and wealthy investors involved in high risk capital investment lose some money in the short term.

”This isn’t unique to cars, every software project looks easy at the start, you can even get “90%” there where it looks and feels like a real solution. But then real life usage testing starts or the implementation process exposes a thousand small problems you didn’t except in the last 10%, or your own original design wasn’t flexible enough for some constraints requiring a significant rethinking of that 90%.”

Yep, and in silicon valley, anyone who seriously plans for the consequences of this is dismissed as insufficiently futurist.

It’s hard to thrive as a responsible engineer when there’s always another huckster around the corner, willing to promise that 5-year future today.

It's like they say: the first 90% of the work takes 90% of the time, and the last 10% of the work takes the other 90% of the time.

Deep learning as an approach has the specific issue that already overconfident people can make shockingly fast progress and get to an 85% solution quickly by investing a lot in low-difficulty labor like data set curation. This lends a crazy amount of misplaced confidence to the people involved as they start trying to get to the last few %.

Most software people have instincts that are all wrong on this because it’s the opposite of the usual experience.

Perhaps someone with more self-driving car knowledge can help enlighten me, but from everything I've read it seems like the biggest problems are from unexpected behaviors in urban environments (e.g. cyclists going the wrong way, stopped delivery vehicles that folks behind must scoot out into the oncoming traffic lane to drive around, police officer directions, etc.) and I just don't see how those problems can be solved without artificial general intelligence, which is a long ways off.

For restricted access roads (i.e. highways), though, the technology seems ready to go. I know there is a lot of concern about the "hand-off problem", but that doesn't really seem like an insurmountable issue. If say 35 mins of a 45 min commute is on a highway, and only the last 10 mins are in a city, that seems like it would be a huge noon to tons of people if the car could take over, completely, on the highway. Are manufacturers likely to take this middle step first?

Yes, see Tesla Autopilot...

Autopilot works but only in a very clean cases highway environment. It cannot with a lot of unexpected situations, like lane closures, stopped cars, and many many others.

It’s good assistant, as those situations happen rarely, but they do happen and autopilot will crash if human won’t supervise and takeover

... run right into the side of an 18 wheeler.

once. they still crash less often than human drivers crash their cars on the highway. I don't want to say Autopilot is good, but everything else is also terrible.

Since Tesla has never released the data for that claim, taking it at face value seems pretty iffy given other claims by the company.

> once.

At least twice.

In conditions suitable for Autopilot in the first place. Autopilot is most often turned off when things get 'messy', a luxury humans don't have, and that Tesla often "neglects" to mention when they're preening over this sound bite.

those are bullshit stats.

They compare overall crashes in all kind of normal highway environment and all kind of drivers and vehicles with the very selective places where the Tesla users decided to enable autopilot. Tesla users are also more educated and have better vehicles which has been shown to reduce the comparative crash rate by a lot.

We don't have data to compare but if I had to take a wild guess on crash rates I am pretty convinced that Tesla crashes more than the equivalent driver on the same stretch of highway as of today.

I didn't ask about advanced cruise control. Autopilot still requires your attention and your hands on the wheel. "Self driving", to really add value, means I need to be able to do something else (read a book, watch a movie) safely.

Yes controlled access highways will be first. GM Super Cruise is already in use on some freeways.

Not sure if anyone can speak to it but looking through recent Glassdoor reviews for Cruise suggests a lot of issues with management from engineers...

Happy to see this, my reaction is that they are serious about pushing the plan forward in the long term; and they are careful about managing the risks. That's the type of innovation I admire and enjoy the most.

Just curious to me how so many "experts" called this a solved problem in the early 2010s, and only just now are realizing the n² * n² * n² edge cases that seem difficult to handle.

Meanwhile Waymo doesn't ask anything to anyone and just do their thing, repeatedly seeing competitors move ahead through PR move before dropping back through reality. If there is one company I would bet money on actually getting there, it's them.

You may be correct, and they do seem to be making real strides. I hope its not because they have been lucky until now. (which statistically speaking is unlikely)

Welcome to the trough of disillusionment.

Quite sad, I worked on one of these cars, and spent late nights getting my part ready on time.

Then I was a GM tariffs sacrifice.

My understanding is that 5G should fix the winter/rain issues, as you can calculate location more precise.

I've always wondered why politicians here in Germany insist that you need 5G for autonomous driving. I thought tech companies just lied to them to get them to subsidize their 5G networks. Are you telling me 5G is actually used for geolocation? If so, I'd love to hear more.

5G hype is being driven by companies in telecommunications and networking. It’s absolute nonsense.

Location tells you where you are. It does not tell you what unexpected obstacles are in your path and how to avoid them safely. The latter seems to be where a lot of self-driving pipe dreams turn into nightmares.

5G won't fix anything. The location precision is worse than what you can usually get with the latest GNSS receivers. And what happens when the 5G tower is down because a construction crew accidentally cut the backhaul fiber?

> And what happens when the 5G tower is down because a construction crew accidentally cut the backhaul fiber?

I assume the cars would stop and we would have to go back to driving them manually for a few hours or days

The goal is eventually to have vehicles without manual controls, so this won't be an option.

Outside of a few limited areas, that goal won't be achieved in our lifetimes.

Apropos of anything else, the tower knows where it is, unless it's mobile. Having no backhaul doesn't mean it can't be broadcasting / used for location information. They also (almost) always come with their own battery backup / generator.

Why would precise location data help with winter/rain issues? Is GM planning to maintain hyper detailed maps of the entire country? That doesn't seem like a workable solution.

To all the naysayers saying, "I told you so!", keep in mind this is a perception problem, not a technology problem. I've been in a self-driving car -- it drove better than just about any Lyft or Uber I've ever been in. They're already safer than human drivers, with fewer accidents per mile than humans.

And if you eliminate Tesla from the category, since they aren't using LIDAR, you can also point out that a self-driving car has never killed a passenger.

The only reason they have to wait is because they know the public expects perfection. The public doesn't understand statistics. If a self-driver kills a person, it would set the entire industry back decades, even though humans kill people with cars every single day.

People just aren't ready to accept that self-driving cars are safer, even if they aren't perfect.

Edits to reply to comments:

To those saying "It wouldn't work in my climate", that's probably true, but they aren't trying to make it work in your climate. All the self-driving companies are trying to release taxi services that work in cities with good weather.

To those saying "But pedestrians are important too", yes I agree with you. I was trying to point out that the fears of being inside a self-driving car are unfounded, because they are very safe. I still think I'd rather be around a bunch of self-driving cars outside too, but yes, they have killed people outside the car.

And on a side note, this is my most controversial comment ever. I've watched the points on this post as I reply elsewhere go from +3 to -3 and everywhere in-between. I think this right here sums up the public perception issues self driving companies have ahead of them.

The industry could help with this if they provided a solution for liability questions. I definitely wouldn’t like to be run over by a self driving car and then having nobody to be held responsible.

The owner of the vehicle would be responsible. Just like any other industrial accident. If you get killed by a garbage compactor that runs on a timer, the machine owner has to deal with that.

>The owner of the vehicle would be responsible.

I think you're misinformed.

Why should I be? I've been informed, by you, that it's a better driver than me, and I have no control over the software.

Let me say, these are not insurmountable problems, by any means. But the government moves slowly.

If I buy a self driving car now, can I buy insurance for it? I doubt it. Maybe the developers of self driving cars should fund such an insurance.

Tesla is planning on selling insurance and I'm guessing it is to cover their self driving features [0].

[0]: https://electrek.co/2019/05/02/details-of-teslas-new-insuran...

Insurance would be attached to the vehicle, and the insurer would know whether or not it's self-driving.

What liability question do you think there is?

Am I buying the insurance, or is the manufacturer/software company?

You are because you own the machine.

Your analogy with industrial facilities falls apart.

If I buy industrial equipment, it comes with a set of safety recommendations, and requirements. On top of those may also be Labor & Industries, or OSHA.

If I fail to follow these, I as the owner am held responsible.

How do I ensure that the SDC follows these? How much control do I have over that, short of "don't buy/drive it"?

If I do follow these, I may also be hit with an insurance claim, but if I can show that the manufacturer didn't account for something I can probably also make a case against them to recoup some of my loss.

Given Tesla, for one egregious example, and their already demonstrated trend of trying to buck any and all responsibility for Autopilot-related collisions, how likely you do you think they are to step up to the plate for "Yes, our car should have determined that there was a pedestrian in the road and stopped, but didn't"?

Because I see it more that they will instead hide behind current software clauses, "not responsible for issues, errors, losses, defects".

> I definitely wouldn’t like to be run over by a self driving car and then having nobody to be held responsible.

If you upgrade to the special China style self driving package, the car will automatically back over any injuried pedestrians, ensuring that the pedestrian will not have to worry about the problem of who to hold responsible.

I think the public will accept a small number of human fatalities, if they arise from things like people crossing the street at night away from an intersection. Most drivers have had close calls like that themselves. What I think will really set self-driving cars back is if they kill someone in a situation where a human driver clearly would not have. If a self-driving car runs over a six year old girl standing in a street because she's wearing an oddly striped dress that the car classifies as a sewer grate - that will be a disaster for the industry.

Again, that's because humans are bad at statistics. Even if it has that odd behavior, what about all the accidents a self-driver avoids that humans can't, because it has 360 degree vision and much faster reflexes? You never hear about those, because they they happen all the time and don't make headlines. But clearly they avoid accidents humans don't, because their accident rate if far fewer per mile.

I'm really curious where you live or work such that your self-driving cars are better than your Uber/Lyft rides and you're observing all these near misses where a real human would have definitely had an accident?

Also, I assume you are also considering all the times that humans entirely avoided dangerous situations that self-driving car wouldn't because they were able to anticipate the actions/movements of (possibly several) other actors/objects and even communicate with these other actors to negotiate intersections?

Do you have a source for the data that says self-driving cars have fewer accidents per mile? Does that include miles driven on closed tracks or is it just orders of magnitude lower than the miles driven by humans and less statistically significant?

I live in Cupertino near Mt. View. I see Waymo and Apple self drivers daily as both a driver and pedestrian. And I see the Cruise and Zoox cars every time I go to SF. I'm interested in the space so I observe how they act.

I see those cars get stuck at intersections with cars honking at them all the time (it feels like a daily occurrence, but that's probably my memory exaggerating the issue). The problem with anecdotal evidence is we don't have the data on how good/bad it really is. I can say as an SF resident that I hate those cars and driving around them is obnoxious because they brake check randomly and are unpredictable.

I think it has less to do with statistics than the feeling/perception of agency. No matter the risks associating with manual driving that are completely out of your control, driving a car can feel relatively safe because it's your own hands on the wheel. Contrast with fears regarding air travel, where one is a mere passenger with no influence over the flight, despite orders of magnitude of greater safety statistically.

Sarah Gailey wrote a Hugo nominated short story (camouflaged as a scholarly publication) about this scenario: https://firesidefiction.com/stet

> And if you eliminate Tesla from the category, since they aren't using LIDAR, you can also point out that a self-driving car has never killed a passenger.

I think the better reason to eliminate Tesla is that Tesla does not yet claim to have a feature complete self-driving car.

> If a self-driver kills a person, it would set the entire industry back decades, even though humans kill people with cars every single day.

In that case, we have already been set back decades by the Uber crash.

> In that case, we have already been set back decades by the Uber crash.

I mean a self-driver without a human safety driver. Many people blame the safety driver for that crash, so it saved the perception (even though it was totally the car's fault).

Oh, I understand now and that is a completely fair point. The safety driver allowed them to shift blame in a way that will be more difficult without them.

I still do not think that such an incident would cost decades, but undoubtedly the cost would be far higher than the Uber case.

Feature complete is a BS term for self driving. Being able to, for example, make a turn on red light, while cool, is far far away from being able to make a red turn reliably in all situations.

It’s like someone claiming they are feature complete to be a composer, because they finished learning music notes.

Feature complete is a necessary but not sufficient precondition for self-driving. Tesla does not claim to have achieved this precondition.

They talk about 1M robotaxis by the end of 2020, and being feature complete by the end of 2019.

Tesla is either lying or has no idea how complex FSD is.

Fair points, but not really relevant to the point that I was making.

>They're already safer than human drivers, with fewer accidents per mile than humans.

No one in the public has enough data to say this conclusively.

>And if you eliminate Tesla from the category, since they aren't using LIDAR, you can also point out that a self-driving car has never killed a passenger.

Not killing pedestrians is just as, if not more important than not killing passengers. After all a pedestrian has no choice in interacting with the self-driving car. The passenger at least made an active choice to put their life in the hands of the technology.

> Not killing pedestrians is just as, if not more important than not killing passengers.

I agree, but my point was that risk of death in the vehicle should not even be a concern for anyone, yet many will say (including other in this very thread) "I would never get inside one of these things".

I might say “I would never get inside one of these things” out of concern for pedestrians as well as myself.

>They're already safer than human drivers, with fewer accidents per mile than humans.

The average driver-less car mile is a lot easier than the average human driven mile (driverless cars have far fewer miles handling inclement weather and rush hour traffic in non-grid cities than humans do) and the human average is dragged down by a small subset of the population that behaves badly.

Driver-less cars are not yet safer than typical drivers and this is before you count all the fender benders the early ones caused by violating social norms by behaving too conservatively in traffic.

This. Driverless cars work on the easiest roads and the mistakes they make are ones that humans wouldn't make. They also currently have humans as a failsafe if they do behave strangely. The stats would be very different if we just sent all the self driving cars out tomorrow without safety drivers.

>And if you eliminate Tesla from the category, since they aren't using LIDAR, you can also point out that a self-driving car has never killed a passenger.

Umm, this is leaving out the very relevant fact that a pedestrian WAS killed near Phoenix by a self driving Uber.

That person wasn't a passenger. My point was that people who are afraid to get in a self driver aren't looking at statistics correctly.

So you're safe if you're inside a self-driving vehicle at all times? That's reassuring.

And also safer outside of one then outside one driven by a human.

>And also safer outside of one then outside one driven by a human.

This is only a baseless opinion.

That's true of regular cars too. If you're on a road you're much safer in a car than outside one.

Maybe the rest of us should ban cars.

Yes, walking on a road is unsafe. Glad we settled that.

Pedestrians are important though. It feels like you are defining an incorrect problem space and claiming perfection but at the same time other thing are important in evaluating their safety. And we often require new technologies to be safer than what they replace.

Pedestrians are important, and are still safer than being around human driven cars. I'm only addressing those who would feel unsafe inside of one.

>Pedestrians are important, and are still safer than being around human driven cars.

Stop making up statistics.

This statement reminds me the scene from the Deathproof: https://youtu.be/hBqzCXlG7Sg?t=120 (warning: violence, language).

No - but this comment is proving the point.

>And also safer outside of one then outside one driven by a human.

Stop making up statistics

Here's the preliminary NTSB report: https://www.ntsb.gov/investigations/AccidentReports/Reports/...

Where did you drive it? On country roads? Roads in bad shape?

> with fewer accidents per mile than humans

This sort of stat is highly misleading. The data is for testing on very safe very good condition roads and/or highways.

I highly suggest this source if you'd like a real look at self-driving car progress and not self-patting press releases from the carmakers or PR churns from the media:

https://ideas.4brad.com/gmcruise-leaks-show-them-way-way-beh...

My favorite was one he had on the security of self-driving cars (which is laughably inadequate):

http://ideas.4brad.com/disconnected-car-right-security-plan-...

Brad Templeton is a former EFF chairman and a former adviser to Google's self-driving car project (from before it was called Waymo).

> Where did you drive it?

Downtown San Francisco, where we had to avoid many obstacles, many people not following standard rules of the road, and major potholes.

let's drive in Hanoi !

https://www.youtube.com/watch?v=gpsybD1MsNQ

>I've been in a self-driving car

How many miles did you do? How many corner cases happened? What was the weather like? And based on this experience you conclude that it's better than humans? How many Uber/Lyft accidents have you been in?

>The only reason they have to wait is because they know the public expects perfection. The public doesn't understand statistics.

It's way more complicated than obtaining perfection. For example, who is liable when a self-driving car crashes?

Has much thought been put into driving in poor road conditions? It seems in the cities self driving cars are pretty much ready, but I wonder about how they will be in bad snow/ice/hail/wind conditions. Humans are pretty bad at these conditions as well. I see a scenario that the self driving cars have X fatality rate which seems very high, while nobody believes the human fatality rate which is even higher.

I agree that we need to be careful about claiming vindication of our pre-existing beliefs on this issue (or any issue!), but your points aren't helpful either:

>I've been in a self-driving car -- it drove better than just about any Lyft or Uber I've ever been in. Th

Experience on an individual trip is not significant evidence on this topic.

>They're already safer than human drivers, with fewer accidents per mile than humans.

Safer in the specific environments they've exposed them too while rounding out the set of conditions, not a general sampling of typical driving.

I definitely think they'll eventually get there, but that's a separate question from whether they're there now.

> Experience on an individual trip is not significant evidence on this topic.

I agree, it was just an anecdote to color my commentary.

> Safer in the specific environments they've exposed them too while rounding out the set of conditions, not a general sampling of typical driving.

And when they are released for general use, they will only be in those specific conditions. Every company pursuing self-driving is looking to launch a taxi service, which would limit the operation to areas where it is safe to do so.

Not everyone lives in a temperate climate with little precipitation, nor in modern cities with wide streets laid out on a regular grid. I seriously doubt self-driving cars are ready for some of the weather conditions where I live. Freezing rain. Snow storms. Roads increasingly narrowing in winter from huge snowbanks. They might be ready for your use case, but they ain't ready for mine.

Every sane self-driving company is looking to launch as a taxi service for this reason. That way they don't have to operate where they don't do well, and then they can steadily expand their region.

Also, I believe that weather is a simpler problem than unpredictable behaviour from other cars, pedestrians, cyclists, delivery vehicles, emergency vehicles, construction, et cetera which occurs everywhere. Of course, it could turn out that both are insurmountable.

How on earth would you know "they're already safer than human drivers, with fewer accidents per mile than humans"?

> They're already safer than human drivers, with fewer accidents per mile than humans.

So are trains. What's your point? This is precisely why people say "I told you so!" - the level of non-exceptions required for them to get right is so numerous that this is exactly why it's a technology problem. Semantics here are important.

This is classic SV think - it worked for me on my shiny paved road from Palo Alto to Redwood City so it must work for the rest of the world. smh

Trains don't drive on roads? And keep in mind that all of these self driving cars will only be in those specific pristine conditions. Every company pursuing self-driving is looking to launch a taxi service, which would limit the operation to areas where it is safe to do so.

> Trains don't drive on roads? And keep in mind that all of these self driving cars will only be in those specific pristine conditions.

That's literally the point I'm trying to make. Trains only riding on set rails is synonymous with an autonomous vehicles ability to drive on a set level of pristine road. Semantics are important here - these companies keep saying "we've got AD coming!" and then keeping throwing an asterix next to it ("note - in select markets only"). Can you see why people roll their eyes?

Absolutely, it's such an incredible bubble of self-delusion that becomes quickly evident to people who try to use SV-produced products outside of the prosperous, no-weather climate, highly connected world of the Bay Area.

I mean, trivial things like Android Auto looks great, until you lose coverage, or your network gets really crappy -- then it starts to behave in ridiculous ways like: try to send a message, and instead of saying "I can't do that right now because X" or "I'll try later" it spins for a bit and then tells you that contact doesn't exist, etc.

If I can't trust SV engineers to deal with common edge cases that result from a place without perfect Internet on a product like _that_ -- why should I trust them to build a car that can handle a climate like mine, a place that has intense and powerful winters that cover all markings on the road, leave black ice patches all over, and have the cars around you using ad-hoc lanes or traffic patterns due to obstruction or severe conditions?

SMH indeed.

They're not trying to handle your climate. They're trying to build a taxi service in San Francisco.

>with fewer accidents per mile than humans.

any actual public data? Because if you mean Tesla then you are wrong because autopilot is not self driving, the driver is supposed to be 100% involved, so from Tesla stats(that are not public) you can only conclude that driver+driver assist could be better then only driver.

>They're already safer than human drivers, with fewer accidents per mile than humans.

There's no way you have enough good data to make this claim.

But yes if we exclude Tesla, ignore non-pax deaths, and vaguely extrapolate from the data released by current players they must be safer!

Reply on news.ycombinator.com