I find it curious that we are going to hold autonomous vehicles to a perhaps impossibly high standard while we also hold people to barely any standards at all.
Think about this: If you are in an advanced autonomous vehicle and you get a warning that conditions are so bad that the sensors can no longer see well enough to drive and that you should take over, should you really take over or should you stop driving?
This is why level five autonomous cars will never exist. It's not a technical problem; it's a political one.
Its easy for someone on hackernews to say "maybe you shouldn't be driving if the car can't"; try saying that to a single mother working as a third-shift nurse, just trying to make ends meet, who's patients will suffer if she doesn't get to work. Try saying that to a technician who needs to get to a downed electric line in the middle of a torrential downpour to restore power to 20,000 people sitting comfortably in their homes. Try saying that to a Marine deployed in Nangarhar, under fire from enemy combatants.
Level Five autonomy could exist, but what we'd end up discovering is that humans let Jesus take the wheel far more often than the engineers in sunny Silicon Valley think. Level Five might mean seeing five feet in front of you and horrible traction, operating on your knowledge of the road, and just driving forward. Until an organization is willing to program a car to still drive forward in those conditions, level five won't happen. Uber practically shut down their entire program after one person died; over 100 people die every day in traditional vehicle accidents.
Which isn't acceptable either. To some degree, people look towards autonomous vehicles as a way of reducing the number of fatalities/injuries on the road. If they're not better than humans & the other benefits don't add up, they won't be accepted.
Based purely on the evidence at hand, our society deems it not only acceptable but not even noteworthy.
The real figure is around 3.2k [1]
This is all US-centric obviously, but my 2c: We can blame individuals for bad behavior, it's harder for us to reason about complex systems breaking and that scares us.
When an accident happens currently, we immediately rush to judgement: You're a bad driver, you were going too fast, there's order in the universe and you made a conscious choice to disrupt that. We'll really reach sometimes too (I've heard this in real life): You had a stroke and lost control of your car? The crash is still your fault, you should've lived a healthier lifestyle. Or it's your spouse's fault, they should've noticed you were acting funny and not let you drive the car.
When people screw up, it's almost always someone's fault even if we have to really reach for it. We rarely consider the system that led to the failure and absolve the individual. When computers screw up, it's unclear who to blame and it's much scarier as a result, so we hold computers to a much higher standard.
Is that true? I just did some very cursory research, but: https://blog.lawinfo.com/2017/09/06/human-error-causes-94-pe...
Where do you live where this is socially acceptable (the blame the weather part, not the drive in adverse conditions part)?
People on the internet love to say you shouldn't drive in inclement conditions, but this ignores reality. Not everyone is lucky enough to work from home. It has snowed 3 times in the last week in Denver. My wife works in healthcare. She has to show up for work unless it is a blizzard.
What are you going to if it is snowing hard with limited visibility when it's time to leave work? Sleep under your desk or deal with a miserable commute that takes 2-3x as long to get home?
Basically when people say the term self-driving car, they mean a human brain like entity driving the car. Humans don't exactly panic and expect a person sitting next to them control a car in case they can't handle it. Sure in those situation a human would get into accidents too. But in case of a machine, such an accident is already programmed to happen. Its like fixed. Which is a very different thing. In fact its even wrong to call it an accident if you know a panic will happen. Its more like programmed manslaughter.
When your engineers say the problem is "too hard to solve" you need to hire better engineers, because those engineers are lying to you. Just because everyone is racing towards a technology and some teams are doing it poorly doesn't mean the problems are insurmountable.
Once I just saw someone coming in a little too hot and was able to pull forward a bit to make room as they were standing on the brakes. Not a big deal.
The lifesaving situation was different though, I was the only car in the right lane and one of those massive F250 trucks was barreling up behind me, not appearing to slow down at all. I was kind of boxed in because the cross street had heavy, fast traffic, and I couldn't pull forward or to the left since there was another car stopped next to me there. So I just quickly scanned the sidewalk to my right to make sure there were no pedestrians, and hopped up the curb to drive on the sidewalk/grass. I wasn't sure if this would work in my tiny car with small wheels, but I figured it would be better than dying.
Worked great. The truck blew through the light and T-boned a Tesla. The guy had 37 empty Coors Light cans scattered around the floor of his truck.
Have any of the self driving car companies even gotten started on the "when should I drive on the sidewalk" subroutine? No, of course not, they are still mired in the bad, early days of the tech where they can't deal with a missing double yellow. It's laughable, the engineers can do much better instead of just repeatedly lowering the bar.
The goofy robot things Waymo is larping at building do not have obvious or well known break down conditions.
Let's pull it back a little and say by 2021, they release level 4.
If any automaker is going to do Level 5 in the snow. It's going to be Volvo. Being a Swedish company (albeit owned by a Chinese firm now), they have a particular track record when it comes to prioritising safety in terrible driving conditions.
I'm pretty sure that they are going to work it out.
[0]: https://www.cnbc.com/2018/06/26/volvo-self-driving-car-sleep...
But they are doing the smart thing:
- they keep these things separate and always under driver's control
- they downplay the significance of these systems and don't market as THE BEST EVER THING ULTRA POWERFUL UNBELIEVABLE BREAKTHTROUGH UNRIVALLED BY COMPETITION ©™ They are simply adding these features one by one to all their line-up
- they have a specific goal called Vision 2020 [1]: "Vision 2020 is about reducing the number of people that die or are seriously injured in road traffic accidents to zero." This is a way more powerful and important goal than the self-serving goal of creating autonomous cars.
- oh. And yes, they work in conditions that are way more different than California. Sweden has such things as constant low sun, snow, rain, sleet, active wildlife etc. etc.
So yes, all things considered I'd root for Volvo, not for Tesla or any other American company.
+ During a dark and wet night a few weeks ago, I used Pilot Assist (lane/speed/steering control) and it genuinely saw the lane lines better than I was able to during major unexpected downpours. I was in control the whole time, but it was a great help along the way and made the drive a lot more comfortable.
- Some of these systems work together rather poorly. For instance, the Pilot Assist system seems to identify white lane lines better than yellow ones, and on US highways (where the left-most edge is often yellow) it will drift left. However, the lane keeping system seems to see the yellow line just fine, so it will vibrate the steering wheel to warn that I'm getting too close, while Pilot Assist continues to insist on moving left. The solution is to get out of the left lane.
But I'm very optimistic and supportive about Volvo's trajectory in this area, to all of your points above.
The Volvo's system is very clearly about 70% there. Sometimes on sharp curves it just misses the lines and drifts into another lane. Sometimes it seems to basically just follow the car in front, which means that it will blindly follow a bad driver in doing stupid things. However, the upside of this is that its weak behavior (and willingness to yield to driver input) makes it so obviously an "assist" system that I would never really put a lot of faith in it, or trust it entirely. Thus to me it's just sort of a better version of the lanekeeping system.
The Tesla, on the other hand, really wants to drive the car. It hates taking feedback from me, and will often boot me out of autopilot mode if I disagree over, say, my position within a lane. It does a much better job seeing lanes and handles lane changes automatically. But: because it's so forceful and doesn't like my input, I find it much more stressful to drive. I don't trust it enough to let it have the absolute control it wants, and find myself missing Volvo's pretty-dumb-but-not-demanding system.
Sweden has such things as constant low sun, snow, rain, sleet, active wildlife etc. etc.
Obviously it helps that they're already a leading truck manufacturer, but it's interesting to note how many other ventures have gone for the "eliminate the taxi sector and reinvent car ownership" over the low hanging fruit for [semi]-autonomous vehicle markets.
[1] http://www.thedrive.com/sheetmetal/17546/euro-ncap-volvo-xc6...
"While IIHS and dozens of other private industry groups around the world have methods and motivations that suit their own subjective purposes, the most objective and accurate independent testing of vehicle safety is currently done by the U.S. government, which found Model S and Model X to be the two cars with the lowest probability of injury of any cars that it has ever tested, making them the safest cars in history.""
Oy. "that suit their own subjective purposes"? Really?
"So, why the difference between the two crash tests? NHTSA's evaluation does not include small overlap front-end collisions, which the IIHS blames for about a quarter of the injuries and fatalities in front-end crashes."
"Ford Mustang hits the wall with 2-star Euro NCAP score [...] Lack of modern safety technology and other flaws mark Ford's sports car down in the latest round of tests. That's the lowest score a brand new vehicle has achieved in quite some time. [...] In NHTSA tests the Mustang scored five starts in every metric."
[1] https://www.evo.co.uk/ford/mustang/18757/ford-mustang-hits-t...
Volvo XC90 is
Teslas are safest for the NHTSA tests, but not the European tests, which, IIRC are better measures of safety
> In short, sensors on autonomous vehicles don’t work well in snow or rain—and that may never change.
With that attitude, you are stuck in the past and don't want to make progress with technology.
The statement you quoted is qualified, not an assertion of impossibility.
Your claim is silly; the alternative to breathless hype isn't Luddism, it's getting educated about a field and proceeding, if possible, in an informed way. Claiming delusional enthusiasm (with a side of hucksterism and a strong odor of financial incentives) is the only way to advance the state of the art encourages the very worst kinds of pseudo-innovation and chicanery.
If acknowledging the difficulty (or impossibility) of something means someone is stuck in the past and is anti-progress, well, then I guess every safety board in existence is anti-progress.
But that's not what's important. What's import is DC, LA, and many other metro cities that can't build more roads, but still need population and economic growth. For these cases, we're looking at something entirely different. If self-driving cars ever come to be a large share of the cars on the road, they become and unignorable option to lift the congestion ceiling on growth.
For major thoroughfares, you're looking at something close to a controlled environment. Even if there's rain or snow, you could have precision car location, relative to static beacons, and relative to other cars. There's no question where the lanes are, because all of these extremely major routes have detailed 3D models that all the algorithms have access to. I'm not saying we have that now, but compared to the other resource expenditures in self-driving technology, it's practically a trivial thing to do.
People don't need a car to drive them anywhere at any time. They need a reasonable option to get to work in the places where the jobs are. I would say that we're losing that right now, and self-driving cars are a solution, rain or no rain.
> People don't need a car to drive them anywhere at any time. They need a reasonable option to get to work in the places where the jobs are.
Mass transit. We know it works, and works well in circumstances you described. We know it increases in effectiveness in response to investment and added resources. Why bet on a long shot, if we know what works?
The requirements seem pretty clear to me. The system has to be on-demand, able to accommodate sprawl, and lose little in terms of average speed compared to old expectations of personal car transit, and most importantly, it can't demand new physical roadway infrastructure.
I see how it's dubious that autonomous vehicles will fit this bill. Or it may be very challenging and take many decades. But let's be honest about the situation - it's the only option on the table.
Mass. Transport.
There will always be a set of conditions that overwhelm your sensors. It sounds like the definition is bad. but we can probably develop an autonomous car that is satisfactory nevertheless.
To be honest, my biggest issue with it is that often those who use it don’t realize that the burden of proof falls on their shoulder. And it just happens that it’s much harder to proof that something is impossible than the opposite. In one case you to prove you’re right once. In the other, you have to exhaust all possibilities.
When I hear an expert say, “...and that may never change,” I interpret it to mean, “I don’t know how to solve it.”
Our eyes are pretty terrible cameras, it's our post processing that makes us able to drive and unless I'm mistaken I'm making relatively straightforward judgements when I'm driving. That processing and logic is really hard right now but I don't see any reason to believe it is impossible.
I feel like at Some point we have to have enough computing power to just stick Multiple Cameras behind glass with wipers and dehumidifiers and alarms if they are obstructed. Basically just humans but better?
My point is I have never heard about self-driving cars having an understanding of crosswinds. I just hope they aren't as oblivious as my friend.
Krafcik's a world leader in the field of autonomous cars, so we have good reason to weight his opinions more heavily than those of random HN users. On the other hand, his "may never change" verdict on sensor tech is vague to the point of uselessness. Whatever happens, he will have had a point.
A contrived example, yes, but the point is there may well be conditions in which, given a wider context, a human would do things a computer can't because it doesn't have the wider context. The thing is these contexts come up all the time in real life. A human can take responsibility, but a computer can't and a human that doesn't in detail know exactly how the computer is designed and will deal with different situations can't make a reasonable judgement about how far to trust the computer in such situations anyway.
and there in lies the problems for companies developing these systems. with current law they would be liable for any negative outcome. logs would be required to be kept and every decision would surely be pulled apart by whatever paid experts each side can find.
so without legislation, legislation the consumer must be very wary about, comes about limiting liability to such systems they will have to side with caution and even then every event is a lawsuit waiting to happen.
finally, would you or anyone you know trust a system which states it can do something you find adversely risky to actually do with you or your loved ones at risk? what would that take? it might simply be a decision of local values
While a computer can't judge that situation and doesn't "care" if it is destroyed it very well can keep driving no matter what.
It's these out of context situations that are my long term issue with losing the steering wheel. A typical example would be arriving at a country fair and parking in a field. I want to park on the side of the field that's currently in sunlight, but over the next hour will be shaded by those trees, near the gate but not too near the toilets, but not next to that patch of wet ground.
What kind of UI does the car need to have for me to specify exactly how and where I want it to park, without me actually driving it? Frankly I think this is level 6+. Level 5 is usually described as a self driving taxi, but family vehicles are used for things no taxi would ever have to deal with.
Growing up in a land with 4-5 months of winter, I was very comfortable in the snow and ice. Still scary as hell at times but the human body is remarkably adaptable to the environment. 30 feet visibility with snow and ice on the roads engages all the senses (except smell). Listening to the sound the tires are making helps know what they’re encountering. Inner ears helping know if a slide is starting. Sight constantly doubting itself and peering into the barely distinguishable road. All the nervous system helping with steering and the pedals. The road has a feel to it in such conditions that any man made system has quite a ways to go before engaging in it.
I do think it can get there. But it may need something better than our current AI tech, as well as some more sensors than lidar.
I agree that I think it'll get there. But, like air travel there are limits and limits ground flights. When we drive a car we don't have a rigid pre-drive checklist, people in ATC watching out for weather and other aircraft, and then a company behind the scenes making the call on risk of life, money and brand reputation. Sure, weather is probably the hardest to deal with in knowing ahead of time if the trip can be completed with 100% success rate. I think the bigger problem is that as cars start to drive themselves, for people with little drive time throwing those people in the mix to take over will result in a very bad outcome. I have over 20 years of driving in ice and snow and still run across situations that are challenging or catch me off guard. But... What about that 17 year old who's car just told them to take over and they've never had to deal with driving in a storm ever? As we stop the task we lose the skill. Ultimately I feel that's the bigger issue holding things back until the manufacturers can get to Level 5.
+1 - And I feel like this is really what Waymo was saying.
Like... cameras behind some kind of glass with special heaters and moving wipe thingies to keep the glass clean? I am sure I have seen something like this working somewhere.
It's always a sad spectacle watching AI fans disparage humans when one looks at the miserable results of the various self-driving efforts. They should at least wait until they have something working decently.
I find it amusing that AI vs Human is actually 'A group of humans working on a technology' vs Humans. If we say AI sucks it actual that we probably aren't good at making certain type of software. Its funny when we talk about AI like it has some Sentience -- when all the onus on improving it and why it sucks is because of humans. Guess someone should start a 'AI shaming' hashtag! :D
There are sensors that can see through snow. They are very expensive, and only in the lab right now, but in the very near future there will (not "may") be processing packages that can fuse optical, radar, and other data into a coherent picture that will allow a computer to see a snow-covered road, through falling snow, better than a human could ever hope to.
I have seen a synthetic aperture radar system that can see through several inches of snow with such a resolution that it can detect paint lines beneath the snow and create a real-time picture of what the road looks like (potholes and all) that a human could never see.
It is too expensive today, but tomorrow?
Not only that, but with gyroscopes, rotation sensors, and real-time communications between vehicles, an automobile that can see through the snow like it is not there that can also talk to other cars will (not can) be able to know ahead of time what road and grip conditions are like based on vehicles that have previously and recently gone through the area. It will be able to adjust its driving to match conditions that are impossible for even the best human driver who has ever existed to detect.
Having 360-degree multi-band RADAR coverage, fused with optical, LiDAR, ultrasonic, and other technologies will result in systems that can: see that it is snowing, see how much snow there is, see the road through the snow, see ice beneath the snow, predict what the road conditions will be in a certain amount of time based on current and forecasted conditions, inform other cars about current conditions, know what conditions are at points on its route in the future and plan accordingly, and ultimately not care that it is snowing so long as it can cope with the conditions.
Today, hooking up some cameras and LIDAR to an Nvidia Jetson doesn't result in a car that can drive in the snow better than a human.
Tomorrow isn't today.
A human driving in heavy rain at night, hunched over the steering wheel squinting out into the distance, radio silenced to minimize distractions, hazard lights on to warn other drivers, never sure of the depth of any puddle that may be on the road, maybe a little tipsy from dinner and tired from a long day's work, will one day be seen as a wildly dangerous and unreasonable risk compared to automated systems with sensor suites that don't care if it is dark and raining and which can map the depth of any standing water with millimeter precision.
Just a matter of time.
They knew EXACTLY what compromises needed to be made regarding self-driving cars all along. Any real product roll-out needs compromises, and this seemingly moralistic apology for layoffs doesn't cut it. They laid off because they wanted to, because their corporate aims didn't align, and that's that.
Autonomous cars can only treat the symptom, not the disease. The real problem is traffic due to suburban sprawl. People need to quit commuting into cities every day, put $50 tolls on the streets during rush hour into major cities, make bus service free.
Then the author wrote a line saying they won't work in the snow or rain (implying any snow or rain at all). To me, that assertion doesn't follow from the quote.
Where did Krafcik say that they won't be able to deal with light to medium rain? Where did he say that they won't be willing to deal with light to medium snow? There are a lot of other articles asserting these things, but few detailed quotes.
https://medium.com/s/story/self-driving-cars-will-always-be-...
> Waymo’s CEO, John Krafcik, has admitted that a self-driving car that can drive in any condition, on any road, without ever needing a human to take control—usually called a “level five” autonomous vehicle—will basically never exist. At the Wall Street Journal’s D.Live conference, Krafcik said that “autonomy will always have constraints.” It will take decades for self-driving cars to become common on roads. Even then, they will not be able to drive at certain times of the year or in all weather conditions. In short, sensors on autonomous vehicles don’t work well in snow or rain—and that may never change.
Or read https://www.spectator.co.uk/2018/07/the-dream-of-driverless-... or any number of "sobering up" articles from last year.