HN is a pretty critical place most of the time but sometimes it's valuable to take a step back. The challenges associated with building a car that can drive itself are pretty insanely difficult (even something as straightforward as "how do we get the training data?") and the fact that we're now at a point where there's a possibility of self-driving cars working in snowy conditions is pretty awesome.
Your comment made me wonder whether i hadn't miss anything : are car actually able to drive themselves in clear condition at the moment ?
I've seen the demos of the start ups in the valley, the local experiments, but i'm still waiting for a single real-world deployment, even limited to local driving, such as city to city taxi driving, or item delivery.
Except for a bunch of small buses that go at walking speed in a very limited area there are no commercial autonomous cars currently deployed.
There are self-driving cars that drive themselves well in clear conditions, but these are the R&D cars with an expensive sensor suite on the roof and a bunch of processing power in the trunk that can't be found in production cars.
I think you're about 12 months behind. The cost of the (LIDAR) sensor suite is down to $5000 from $50-100k and poised to go lower, and the processing power needed in the trunk is being achived with simple dev boards that cost as much as an enthusiast GPU. You can buy an off-the-shelf Vega GPU and have tens of teraflops on your desk.
That's still more money than not having those components in your vehicle, for sure, but that's well within luxury car territory and costs are still coming down rapidly.
A car that can do a single drive is one thing. A drive that reaches a level where millions of people can do millions of drives with no issues, that's an entirely different problem. We've seen many many videos of "single drives" working well.
Demanding a probability of zero can make anything impossible. And you're not even asking for zero fatalities, you're asking for zero "issues", whatever those may be. That is a wildly unrealistic demand. Human drivers are not that reliable either, especially in bad whether conditions.
I'm not asking for zero, but just saying that scale matters more than most think. Even something like 99% success rate, a video of a single drive would look great, but at scale, it would absolutely fail. 1% is a huge failure rate considering millions of drives happen every single day. That's 10000 failure per day on 1 million drives.
Do you intend to drive on those other days, without any fair weather routine? Automation unlearning is a problem with airline pilots, despite all efforts like continuous flight reviews that drivers don't have.
That is a reasonable question, though we already have the situation where people are driving in snow with little current, or even previous, experience of those conditions. The biggest problem is going to be with people who have never driven.
> The biggest problem is going to be with people who have never driven.
Isn't that basically what @usrusr was refering to? People who rely on the autonomous driving so much that they basically forgot how to drive themselves?
It's ok, context matters. There are places were you won't get snow - ever. So it's not something you'd worry about when deciding if the car is self-driving.
The skeptic in me looks at the current climate situation and recoils in horror at the thought of cars without the 'snow firmware' trying to deal with unexpected snow in South Africa
Hypothetical test: You and a Waymo car try to drive to the main road from my parents' house on a random January morning. It's a harsh test, but no harsher on you than on the self-driving car... so if you're hesitant, are you hesitant about yourself as a driver, or just about the car?
I build autonomous/ADAS testbed cars in Detroit and it's astonishing the stuff that the California cars get flustered by. LIDAR is baffled by snow, RADAR sensors get packed with slush if they're mounted too low, etc.
I expect traction to eventually be one of their strong suits. Machine learning on weather conditions and traffic patterns should work. I’ve got a unique amount of snow-driving experience in most every region of the globe (many years working for a snowboarding magazine) and my hunch is a computer could do this wonderfully. Much of these things should be very predictable with delicate sensory input, and roads have predictable problem spots.
One particularly problematic variable that will undermine weather calculations and traction will be the complications of road de-icer. In high-altitude inland regions, de-icer is not used, with a few exceptions after storms are long gone and the temperature has warmed. These regions should be a piece of cake for autonomous vehicles. Detroit in mid-winter probably resembles this for a few months due to low temps but humidity will destabilize de-icer during early and later winter. And, temp fluctuations resulting from humidity, coupled with high snowfall, necessitates the use of de-icer to prevent the city from getting buried in. In coastal regions like CA, de-icer more predictably melts and cleans the roads. In regions like Chicago, de-icer will hve every effect imaginable, both creating ice and melting it multiple times a day. But if a road is not de-iced the traction will follow a different pattern. I can’t imagine how a self-driving car could determine this. To further complicat things, de-icer can often create ice and be counterintuitive in only shaded parts of the road. De-icer is used so differently, an so often foolishly in different regions.
Detroit doesn't really get enough snow or stay cold enough for ice to be an extended problem. On a cold morning intersections can get a bit slick, but that is about it. (I've lived in Ann Arbor, my brother lives a bit outside Detroit).
Michigan also tends to just salt the shit out of roads. Probably enough to impact the chemistry of the Flint River. Not so much way up North where it stays colder. Sand there.
The cars certainly don’t measure the traction, not even once per second. You need wheel slip to do that, and constant wheel slip is not a good driving mode. Not to mention the difference between static and dynamic friction.
At most cars can collect the lower bound of friction coefficient when braking/accelerating and lower/upper bounds when wheels are slipping. If you have a straight stretch of road, and no cars tried to brake/accelerate on that, the only friction data you would have is ‘no wheel slip at .02g acceleration/deceleration’. It doesn’t help the AI predict its braking distance for an emergency braking maneuver when driving that stretch of road, so a conservative AI would be forced to keep a much larger headway than a typical (or even cautious) human.
Have you been here long? We call it a “Michigan winter” for a reason. Yes, Detroit suffers less for a few reasons, but I still wouldn’t agree with your assertion.
It doesn’t sound like you’re describing the Detroit I grew up in, where dozen+ car pileups are a regular occurrence, particularly the big hundred+ ones on 94 out your way. That’s even discounting the wide-area ice storms that knock out power from everything getting coated in three to four inches of pure ice, which I can remember happening at least four or five times when I was younger. The last couple winters have been relatively light. We are due for a whopper.
Seriously, Michigan winter is accelerated natural selection for drivers. You learn quick or you die.
Ah, I thought Detroit would be colder than that mid-winter.
As for traction on ice, if the vehicle is already on the ice, it’s too late to slow down. I wonder how the cars could know when they are coming up to ice. Humans can rarely know, but have some complicated ways of identifying, I think from glossiness? And ice is much more rare than snow/slush but both can be present side by side. Either way, I think they will figure it out. Sand can be awful to clean up in urban areas when too muh is used.
It's a mystery to me why people assume that California car companies don't test their cars in Tahoe, which has snow. Where we go skiing. Does Detroit not have areas within driving distance that have radically different climates? California does.
The article even says it: "But it won’t be Waymo’s first foray on dangerously snowy streets: the company has previously tested its vehicles in winter conditions outside Lake Tahoe."
I wouldn't be that optimistic.. I grew up in Duluth, MN. A city that in 2 miles goes from 600' ASL to 1400' ASL, there are some roads with a greater than 10% grade to them. It snows there, and not just any snow, lake-effect snow. You really do need a special set of skills and roadway awareness to navigate inclement weather in cities like this.
Detroit is a great step, but I'd really like to see some more challenging areas handled before I'm convinced.
I think a good question that follows from that is: how unsafe is too unsafe for autonomous vehicles? Can we evaluate roads in some generalized way that will not only prevent autonomous vehicle activity, but inform local governments about when they should institute driving bans?
In an ideal world, I trust autonomous cars over human-operated cars. So if the conditions are too dire for autonomous cars, I really don't want to be out there with humans driving.
> but inform local governments about when they should institute driving bans?
In the U.S. I'm not even sure what such a mechanism would look like? How could you enforce this?
> In an ideal world, I trust autonomous cars over human-operated cars.
I don't understand this mentality out-of-hand, especially because I've been in cars that have had mechanical failures. There is no mechanical system that is perfectly reliable, and I'm not sure we're considering the emergent effects of this properly.
Finally, unless you outright _force_ all drivers to use the autonomous mode, then users can still take control of their cars as they see fit and we're really not improving anything. Case in point, the video says that driving deaths disproportionately affects young _people_. Which is garbage.. it affects young _men_ who are 8 times more likely to die in a vehicle that they're driving than a female of the same age. I don't know how autonomous modes being available are really going to help this group.
> Finally, unless you outright _force_ all drivers to use the autonomous mode, then users can still take control of their cars as they see fit and we're really not improving anything.
I mean, Google has prototypes with no steering wheels, and has as an explicit goal the ability to transport children, blind people, the elderly, etc. Eventually they absolutely do want to force drivers to use autonomous mode (at least "force" in the sense that users will need to opt in when they choose a vehicle or service, rather than minute by minute during a trip).
If it is possible to get around in a car without driving, the standards for licensing can be increased.
So all those young men can still do all the things they need a car for, they just don't get to legally operate one until they have demonstrated a higher level of proficiency.
And all the ones that aren't just total assholes can let the car drive when they have been drinking.
Do things like speed and drive dangerously? I'm sure they'll pass a higher level as well. Unless you'd implement a psychological test that'd reject too temperament people.
You don't necessarily need psychological testing. For instance, just make it easier to lose a license. Minor speeding could come with a license suspension if such a penalty were less consequential.
Over there "dangerous driving" means license suspension for up to 3 years. Doing donuts, speeding and overtaking recklessly, running lights while endangering oncoming traffic.. Any of these are immediate suspension and hefty fine. +50km/h (or +30km/h for novices) is immediate suspension as well. We still got plenty of jackasses. Mostly young men. With or without licenses...
Yes of course, but I'm talking about a 6 month license suspension if you get a ticket for going 7 over, not a suspension after repeated severe tickets.
Many states automatically revoke your license if you get convicted of any moving violation under the age of 18.
IMO it's a dumb, inefficient drag net because being black and driving 5-over is not a strong predictor of whether or not someone will drunkenly drive off a cliff on their way to prom.
Roads get shut down all the time in inclement weather, even interstate highways. There are also things like 'chain advisories' which prohibit traffic of vehicles unequipped with tire chains.
What this looks like is generally highway patrol blocking off the road and the news of the road closures or advisories is broadcast on the radio and television (and now the Internet, I suppose.)
Even down in sunny Austin Texas we have road closures when heavy rains inundate low-water crossings, though it seems every few years someone tries to make it through one and ends up drowning.
> In the U.S. I'm not even sure what such a mechanism would look like? How could you enforce this?
Some entity produces the data that finds roads unfit for driving based on weather and publishes it, local governments consume that data and institute a driving ban, non-essential personnel cannot drive or they'll get a citation.
> I don't understand this mentality out-of-hand, especially because I've been in cars that have had mechanical failures. There is no mechanical system that is perfectly reliable, and I'm not sure we're considering the emergent effects of this properly.
Correct, but I believe that a well-conceived autonomous car has a much higher standard of proper operation than a human, so even if there were mechanical problems, I'd trust the autonomous car to halt operation before the human.
I also don't really understand your last point so sorry for not addressing it.
Ahh, but you are assuming failure. What if the human has a stroke?
Really, you must compare failure rates and severity on both sides. Looking at say aircraft, autopilot cause fewer crashes than pilots and spend vastly more time flying aircraft.
We have been building cars for quite some time now, and yet there are still millions of recalls year after year. The idea that IT can solve physical problems is so pervasive on this site.
> autopilot cause fewer crashes than pilots
How often do pilots really cause crashes, though? In the sense that it's the pilots fault-more often than not, it isn't. Most crashes are not the result of pilot failure in normal operation-but pilot failure in situations out of their control.
It's hard to blame a pilot when an engine cuts out (a mechanical failure).
my point is not that automated cars are never going to work. But it's just a utopia on here, about how all of our problems will be solved. Computers can't beat physics.
> young _men_ who are 8 times more likely to die in a vehicle that they're driving than a female of the same age
To paraphrase an acquaintance who drives a tow truck:
Basically all deadly crashes are a result of an inability to drive properly (either via distraction or impairment). Actual "people being bad at driving" rarely causes deadly crashes by comparison. People make bad decisions that make them bad at driving.
When a drunk couple goes off a cliff it's always a dude driving. Dumb things involving cell phones are less deadly but skew more female.
It’s also possibly not worth solving. I don’t know off the top of my head but I’d assume most of the population lives in temperate cities that rarely see the kind of icy weather regularly. Your vehicle could just not drive automatically in suboptimal settings and I think that’d be totally fine.
We allow humans that have never driven in adverse conditions before to hop in their cars and take a road trip to Duluth. If we plan to ban robots from doing similar things then maybe we shouldn't allow humans either.
Is there progress being made on letting cars self-drive on unmarked and dirt roads? I imagine that would be important for users in less developed areas.
Sounds like it would be a good step to take before taking on snowy roads. The conditions are pretty similar with low traction, no markings, People driving in opposite directions having to figure out a way around eachother when the road narrows... At least you wouldn't have to deal with the actual cold temps and snow and ice. Traction would be more consistent on gravel vs snowy and icy roads too.
The video in the top voted sub comment (now) shows the Tesla unable to navigate a two lane suburban road safely for long with the youtube comments/excuses saying it's an early release of non highway software (2017). I don't know anything about it technically but that is not particularly confidence inspiring.
Well-graded dirt roads in good weather aren't really challenging other than not having lane markers. (Which is pretty much a trivial issue for a human.) Maybe you need to slow down a bit if it gets washboardy.
The challenge is as roads get narrower (<2 cars), rougher, etc. to the point where you really need high-clearance 4WD and may even have to deal with washouts, etc. I don't expect you'll see autonomous driving systems get here for many decades if only because it's not worthwhile to do so and so much is dependent on evaluating the current conditions on the ground.
Just an aside but you actually want to do the counterintuitive thing and speed up on washboardy roads (provided it is safe to do so). Higher speed means your vehicle won't have time to be pulled down into dips letting you ride smoothly supported by the peaks.
That's because the best way to drive in the snow is to not do it, yet that doesn't stop plenty of sport compact cars without properly rated tires trying to take on steep grades in sub-zero temperatures while its snowing sideways. There's no skill that's bailing that person out.
Ah people love to hate "sport compacts." Really, the worst offenders are SUV drivers with AWD who think that they can take their bald-ass no-season tires and get anywhere just because they have the appearance of ground clearance and an extra traction wheel. Last time we had a major snow, I took my "sport compact" car with proper tires and ran circles around the clueless SUV drivers who were spun out in the middle of the highway.
The best way to drive in the snow is to get out and do it at a time when there is light traffic. Go find an empty parking lot. Learn how your car handles, learn to rock your car to get unstuck. Learn about gearing, learn what no traction feels like. Learn what a skid feels like. One day, you might not have a choice but to drive in the snow, so you might as well spend an hour or so working on your driving skills. Not to mention all these skills translate to making you a better driver all the time, and could help you some day avoid an accident.
I mentioned SCC because I used to drive autocross with an SCC and cared after my cars. In reality it's everyone, as you noticed, but plenty of my dumb friends were running their customized SCC cars into snow banks due to Z-rated tires on during snow season.
I was driving my Subaru through Idaho on I15 during the winter. Going down one pass, I put my car in first gear and prayed because even though I had AWD, no snow tires or chains meant I was in for a hard time.
> The best way to drive in the snow is to get out and do it
In Sweden, before you can get your driving license you have to take a course where you spend an hour on a skid pan, taking corners too fast and emergency braking (a "moose" shows up out of nowhere) so you know how it feels and what kinds of distances and reaction periods are involved.
Also in winter, snow tires are legally required between december and march (and this is enforced with random DUI-checkpoint-style stops)
Winter tires are not expensive at all, and anyone who can afford a car should be able to afford one. You can get 4 Continental (best in the market) for less than $200 (15 inches size).
>and anyone who can afford a car should be able to afford one.
Allow me to add the obvious implication
>anyone who can afford what I consider to be the minimum car worth buying should be able to afford one
Plenty of people drive cheap junk they bought with their tax return and pay insurance month to month. Those people deserve to get to work and the grocery store without having to violate any laws. The damage to our society from not requiring people to use snow tires is far less than the damage to our society that would be done by a huge chunk of people in violation of a law for a good chunk of the year.
Laws are how we quickly and easily categorize what a society finds acceptable and unacceptable.
Laws are not how you determine what is and isn't acceptable.
Our society does not require snow tires for part of the year. If this is so abhorrent then write whatever elected official represents you or move to northern Europe.
> That's because the best way to drive in the snow is to not do it
Good luck with that if you live in a northern climate. Modern life would essentially stop functioning here in Minnesota during the middle of winter if no one drove because of snow or ice.
Between the efforts of the DOT (plowing, deicing) and following some common sense (slow down, increase your following distance, avoid sudden changes in direction, anticipate stops, ensure your lights are not obscured by snow etc) there are only a handful of days when travel is not advised (usually due to blowing snow).
In part because I mostly don't commute these days and generally have a fairly flexible schedule around most things, I've been able to mostly avoid going out in snow in all but the rarest circumstances. I don't miss it.
Ignore the speed limit. Go however fast is reasonable for the conditions.
Going 35 on a road that's marked 45 but usually driven at 60 in good weather isn't going to give you time to stop short for something so you might as well go 45 if that's reasonable and piss off an order of magnitude less people.
Source: New Englander sick of being stuck behind people like you.
Could the camera use the clear view screen [1] instead of counting on windshield and wipers? Maybe it does already, I don't know.
Clear view screen is a round window that is usually located on a normal window. This window rotates at ~1500 RPM, thanks to that it will not be covered by snow or rain. It is used on ships and in locomotives.
One of the biggest challenges with self-driving cars in snow is the snow in the environment -- for example, how do you see the lane markers so that you stay in your lane?
Like others have said, the vehicle will localize using GPS, LIDAR, other image features and then attempt to continue in it’s route. If the confidence is low the system will most likely just take no action.
I did the self driving car class on Udacity and it was super effective as a primer for how these systems work in general and framework for solving these problems.
This is also a human challenge. You follow the tire tracks. If there are no tire tracks, guess. You can have a margin of error of half a lane and be within human parameters.
How do you guess? Oh, you need another set of hundred million miles of testdata for this case?
The thing is that it is vastly different from any other scenario a self-driving car will ever experience or train for. Humans on the other hand grow up with snow and know how it behaves, how it covers the environment and what features are prevalent despite of it (and consequently, people not used to snow that suddenly finds themselves in a snow-covered landscape often fail miserably).
Falling snow can significantly cover both video and disrupt LIDAR immensely. Just as our vision, thing is we already know how to cope with it, it is a greatly different problem than driving in good/ok conditions.
I imagine they'll just have to do what humans do. Follow existing tire tracks if available. If not, just intuit where the road using your best judgement. It's not terribly unlike driving on dirt roads (though harder, surely) which autonomous cars already did surprisingly well in the first two Grand Challenges (no lane markers there either).
> for example, how do you see the lane markers so that you stay in your lane?
As a human in a snowy region, who has been in some pretty dicey conditions, I measure the slope of the road. Roads are typically shaped like a triangle, so if your car changes it’s rotation on the axis that is your direction of travel, you know you are outside of your lane. Even my smartphone has sensors for that kind of measurement.
I wonder if it would be workable to have a half-automated mode for these conditions...the human steers, and the car does all of the rest (accelerating, decelerating, etc). Or the human could have a binary go now/stop now control, with the software handling the details.
Detroit still seems like an easier environment than Uber/CMU's Pittsburgh, home of barely-maintained brick roadways with blind corners on 20%+ grades and icy winters.
I'd have just let this joke slide if Pittsburgh wasn't a pivotal location in radio history due to KDKA. :P So let me just say 105.9 The X for life (as long as Mark Madden isn't bloviating, anyway).
Still very common elsewhere. As noted by the article this is a widespread phenomenon in the Northeast... I see it all the time in Boston and Worcester.
Pretty much anything in the US is easy. Try Indian cities, Italian villages or the Russian countryside. I’m absolutely certain that not a single one of the so called “self-driving” cars could handle even the most simple challenges in these environments.
I bet self driving cars are first adopted enmasse in Asian cities (primarily Chinese ones) given need. Even though the driving (and pedestrian) habits seem to imply otherwise, the chance to optimize limited road infrastructure will just be too juicy for the politicians to pass up. In fact, lots of companies are working on self driving cars in china right now for this reason (also, it’s much more likely given china’s autocratic government).
The US is pretty big. I have neighbors that ATV and snowmobile the last few miles to their house, except during the very few months that they can wrestle an old Jeep up there. Google Maps isn't very helpful.
There are very few brick roads in Pittsburgh. Most are in older sections of the city where the residents have lobbied against a paved road for one reason or another.
And I would imagine steep hills and icy conditions are exactly what you want to test self-driving cars on to see how they handle it.
To be pedantic, I think saying there are very few is misleading. There are very few outside of older neighborhoods like you say, but there are plenty of small brick/cobblestone roads around. Just not many heavily traveled or particularly long ones.
> There are very few brick roads in Pittsburgh. Most are in older sections of the city where the residents have lobbied against a paved road for one reason or another.
Many places still have brick. It's not a majority, but it's not a couple of historic district streets either.
I'm not an expert in perception or autonomy, but I've always had a loose sense that winter driving would be an order of magnitude more complicated for autonomous vehicles. It had me worried: what if technology presents a future where the world is split up into industry regions based on climate? You have some serious costs associated with doing business in a region that has winter.
"Available in Canada (summer only)"
"Shipping to winter regions is far more expensive because we need to use human drivers half the year. So we generally avoid focusing on them."
Not a new problem. Geography has always strongly impacted industry. But interesting to think of it from this perspective.
I'm excited to see attempts to tackle the truly difficult form of autonomous driving.
The rules for driving in snow are pretty simple. Go slow, pump the brakes, give a bigger safety cushion for obstacles.
The problem is likely around vision. A lot of winter driving is done at an unsafe level of visibility.
It is probably better having non static rules for roads. Snow should reduce speed limits, and white out conditions should be used for emergency travel only.
Unless you can utilize radar to obtain better visibility than human drivers, in which case you could probably drastically reduce the number of winter accidents.
And heavy rain, heavy rain at night with street lights reflecting off the pavement right into the cameras, black ice, fog, giant snowflakes, high wind snowstorm, etc etc.
A cop wearing a parka directing traffic in a snow storm at night.
And what about criminals tricking cars into doing things?
It will be a long time before smart people trust automated vehicles to take their kids without adult supervision. Cars will have to be able to interpret our actions and intent better than we can. Or at least be able to defend their cargo, record evidence and alert authorities.
It's not really a normal place to be testing in traffic in many ways. During the daytime and, sure, I guess self driving cars can get used to the absence of left turns on highways and the 'turn arounds'!
At night, I was told by friends who are locals when staying there, as have other visitors I've talked to, that you do not want to stop at traffic lights in many areas of Detroit at night, at all. One person told me he was pulled over while stopped at a light and told this by a police officer, even.
Apparently, in these fairly deserted semi-industrial areas, stopping at a red light simply makes you a target. I was also instructed to leave room in front of me to move my car if someone tried to enter my car at a traffic light, and to be ready to roll up my window on someone's arm if needed.
Getting tech employees off cars doesn't need self-driving cars. Google is perfectly happy to have shuttle buses, and pay for them, provided that housing is sufficiently dense, and there is a place where the buses can legally pick up and drop off employees.
Probably the biggest problem with using mass transit is latency; having to wait around doing nothing until your transport arrives. It's a constraint on your schedule and adds a certain amount of stress to your day. And if your desired departure or arrival time don't coincide closely with the schedule, you're going to waste a bunch of time even if the transport is quick.
Subways solve this problem by running every couple of minutes. Buses generally don't solve this problem well because traffic and alighting / loading delays cause them to bunch up. But on-demand self-driving cars can solve this problem, as long as there are enough of them close enough.
Having ridden the Google buses for 6 years, I can tell you this generally wasn't a problem. They were almost always on schedule (I'd say 95% with 2-3 minutes of planned departure) and never bunched up. You didn't really have a latency problem, because you knew exactly when you needed to be at the stop, and the bus would be there right on time.
One bus replaces probably 150 car trips a day (assuming it makes ~3 runs, partially full. If you assume similar capacity utilization on cars (3 passengers), it would take ~50 car trips to replace those 3 runs, effectively putting 17 cars on the road for every bus currently in operation.
Bay Area buses don't solve this problem but many bus systems do. Compared to an effective bus network, Bay Area buses suffer from:
* Short distance between stops (due to low housing density).
* Traffic due to lack of bus lanes.
* Long time spent at stops due to cash payments and/or single door entry, rather than proof of payment or tap on stations at every door, no cash allowed.
* Inaccurate/nonexistent realtime updates.
A bus every 30 minutes isn't a big deal compared to SF/Bay Area travel times, but only if I can rely on a timetable/GPS to arrive at the stop 2 minutes before departure and be confident with a scheduled arrival 2 minutes before my appointment.
Low density means that there are less people per km within walking distance of a bus stop which means you pick up less people. So it’s less efficient in that way. High density means at each stop there is much higher amount of people who live near the stop.
How do you think that means shorter distances between stops for lower density? You pick up less people per stop, but it's less also people over a given distance, so it doesn't imply stopping more frequently.
"Short distance between stops (due to low housing density)"
If anything, you have it backwards. In most of SF, places with high housing density (e.g. Van Ness) have their stops closer together than places with far lower density (e.g. Geary in the outer Richmond).
I suspect (but do not know for sure) that this is done on the theory that dense neighborhoods require more granular transit access.
I have noticed that waiting around on the subway platform or at the bus stop is way more stressful than sitting on the subway or the bus, even if the subway/bus is just stuck for some incomprehensible reason and showing no sign of progress towards my destination.
This leads me to suspect that the stress of waiting for public transit is largely incidental and could be reduced by better station design, without having to make any (difficult, expensive) improvements to service frequency and reliability.
> Subways solve this problem by running every couple of minutes.
Similarly, in many transport corridors intercity rail runs several times an hour (not so much in the US) and shuttle flights run every hour or two. Which are also sufficient in those contexts, in that it becomes reasonable to show up at the station/airport without checking the timetable.
more housing = more residents = more local tax income
That's the formula San Jose followed for two generations and is exactly why it's less successful than its neighbors, who pursued industrial development plus "big box" retail (both of which provide a better local tax base). Sunnyvale, Mountain View, etc. used San Jose as a "bedroom community" and instead welcomed a more industrial base.
This sounds like a nightmare. Employees who would have walked, biked, or bused would now be added to traffic. It sounds way more convenient to be picked up by a company car than to bike, especially if it’s raining.
Google having full parking lots doesn’t harm Mountain View and Sunnyvale traffic as much as extra trips do.
I'd much rather they figure out snow conditions. Driving in a snowstorm is challenging even for experienced human drivers, and I'd feel far safer in a autonomous car that has been shown to handle snow well than one that merely dogfoods with employees in sunny California.
Because snow conditions are inherently more dangerous and more likely to yield data for the cases we really care about, i.e. accidents and near accidents. Getting commute data in California is likely to just give you a ton of boring repeat data going to and from googleplex and a few locations. As far as I understand, the technology is already good enough to handle that sort of 80% reliability stuff. It's the last 20% - the hard stuff - that needs to be hammered on from now on.
Someone had a good example in this thread, for example, you're more likely going to run into issues with clogged sensors in detroit than you are in palo alto. Another thing that happens way more in snow is automatic torque rebalancing in 4WD systems. Rainy days aren't nearly as common in Palo Alto as snowy days are in Detroit but both share the property of poorer visibility and an increase in pedestrian distraction levels. Etc, etc.
A lot of technology only exists because their original problem set was extreme in nature (e.g. things out of NASA or military research). I think a lot of "easier" problems in autonomous driving could benefit from a breakthrough solution that got discovered in the course of solving a harder problem.
Personally, I'd like to see these self driving cars routinely tackle more challenging climates than those that you can suntan in for the majority of the year. The snow covered streets of Detroit should yield some interesting data.
I feel like this is too short sighted. That will inherently happen if they are able to tackle different terrain in different weather conditions. Google has offices in more than just Silicon Valley.
As a society, are we ready for self driving cars? It seems to signal the start of the next industrial Revolution, with truckers and taxi drivers going extinct, with more industries to follow. Machine learning has so far been kind of in the background, with this it's very front and center and will directly impact millions of people.
Also, I've always wondered, if self driving car crashes who is legally liable? Is it the programmer or the occupant?
The problem with self driving cars is that you can never rely on a human to take over at a moments notice. Google perfected the self driving car over a decade ago but found that once on the motorway (Freeway for my transatlantic cousins) they literally switched of, rummaged around in the back did everything other than be able to take over the wheel at a moments notice. So google had to go back to the drawing board and design a car that could drive literally everywhere! We covered the topic in part of our recent podcast. If you'd like to listen debate or disagree! Skip ahead to 23.03 minutes.
Is Waymo still planning to launch soon in Phoenix? They were going to start offering rides to the public this fall without the employee in the car. Or did they change their mind? If anyone has ridden in those vans I am curious to know how well their early rider program is performing in terms of not needing the human driver to intervene. I check the news for a Waymo launch every day, because it will be one of the most significant events in the history of self driving cars.
> Starting in April, Waymo began testing a free service with select passengers in Phoenix, using some of its 600 minivans from Fiat Chrysler Automobiles NV.
Yeah that's old news. I am referring to recent articles from a few weeks ago where someone suggested that there was at least a rumor that they might launch without any actual driver to the general public in Phoenix this fall sometime.
The rumour traces back to a single anonymous source reported on by the Information. I hope it turns out to be true, because in 2012 I made a 5 year prediction for it and it's always nice to be correct about things, but we'll have to see. From the outside looking in, they look to be pretty close, but only the people at Waymo really know how close they are to doing it.
But right now there is a driver in the car ready to take over. If they can do it without drivers being in the car, then driverless cars are just around the corner. I'm skeptical.
I'm curious to see if we'll see the Dubai style autonomous drone taxi service thing come to the US at some point. At least in major cities. Personally I'm in favour of less cars and more walkable/bikeable urban spaces. https://www.lonelyplanet.com/news/2017/10/03/flying-taxis-du...
> "Having lived through fourteen Michigan winters, I’m confident that there are few better places that will prepare our self-driving cars for winter conditions"
Lived in Michigan for 30+ years myself, on the West side of the state. We get a lot of lake effect, and it's only 3 hours from Detroit. It's also 6 hours from the middle of the U.P., where they get _real_ snow. And if they really want to, they can go into Canada.
Michigan is a great location to test self driving cars, because the lake effect makes the temperature fluctuate around freezing. This means we go from icy to wet frequently. It also helps that we get a decent amount of snow, giving us a much wider range of driving conditions than the Bay Area.
Context is a major part of driving it worries me when you see these videos of self-driving cars slamming on brakes a metre or two from a pedestrian. A human driver would see the person at the side of the road whether it's a marked crosswalk or not. There are visual cues we humans know which mean "I'm going to cross the street now" such as a pedestrian turning their head towards the driver to make eye contact first, holding out one arm while looking both ways. Slamming on the brakes at the very last moment isn't what I'd call a success.
As a Canadian who lives in the very snowy Maritimes I have a lot of experience with driving in the winter. I started driving at age 14 and my job as a teen, 20s and 30s required me to drive in all conditions. Winter driving has to be instinctive there is no time to think and as noted above also depends on context.
If a pedestrian walks out from the side of the road and that road has a curve maybe even a downhill curve what now? Slamming on the brakes even ABS may not help. The inertia and being on an angle may (my guess is will) cause the vehicle to slip towards the outside part of the curve. If the vehicle is in the right lane it would slide into oncoming traffic.
Counter-steering at the very instant your vehicle starts to fishtail is a learned skill that has to be instinctive. And when to use that technique or when not to use it.
Slush is I think one of the most dangerous things a winter driver will face. A driver who drives into a line of slush that may be left on the center line can easily have their vehicle jerk violently to the left. It can be a situation where you have less than half a second to react.
Brakes on a curve has to be one of the most common problems with inexperienced winter drivers. I see it all the time even with drivers who I know have been living here all their lives.
Who will scrape off all the optical sensors in the morning? I've had ice so thick and so hard to get off you'd swear it was concrete glued to your windshield. Not to mention the slush, snow, dirt-salt-water mix constantly splashing up. Anyone with a backup camera knows it's a constant battle to keep it clean or it's useless.
It will be interesting to see the results but I think there is far more to self-driving in the winter than on clear, dry roads in warmer climates.
I lived in Downtown Detroit for a two years. Unless things have changed since 2011, there aren't many streets outside of a narrow strip of Mid Town where the city actually plows the streets. I'm curious how examples of government mismanagement like this get highlighted. I can see the headline in the Detroit Free Press now "Snow No-Go Zone: Waymo won't drive East Jefferson until the city plows it."
Exciting. Test it in rainy Seattle and freezing Alaska, too. If this works, this would show real, tangible leadership in the field. No one can do this yet, and autonomous driving tech is not to useful if it craps out in inclement weather.
Seattle can have "grey-out" conditions, where spray from other cars combines with the rainfall to dramatically limit visibility. It's rather terrifying to drive through, where you can only see for about 20 feet in any direction. Autonomous driving that entirely relies on computer vision will fail just as poorly as human eyes - multiple sensor types are necessary. How many types of sensor does this car have? I wish the article would have gone into that detail.
Why would we expect computer vision to perform as poorly as human eyes? The input sensor is just part of the system - human eyes aren't very good sensors (narrow FOV, narrow set of usable wavelengths, etc) and the human brain isn't very good at detecting things (gets distracted, fooled, etc.) Anyone who's been using Google Photos search will tell you if often notices objects in the image that you didn't even register when you took the photo or viewed the image.
I'd expect the opposite - a self-driving car might outperform human drivers significantly in conditions where we tend to underperform - poor visibility, longer braking distances, poor steering control - because it will be more effective at adjusting than a human.
And maybe some kind of communication/message passing between vehicles maybe useful. For example, if a vehicle senses black ice in a certain part of the road, it can pass that information to other cars so they know to slow down when going over the same part of the road. So something like Waze but for road conditions.
> And maybe some kind of communication/message passing between vehicles maybe useful.
I've made this suggestion before, but all commenters remarked about the risks of trusting another car, trusting another signal, hackers, script kiddies, and other malicious agents. The safest inter-car communication was determined to be adaptive cruise control.
We already trust our safety to the cars in front of us - working tail lights & signals, proper securing of rooftop loads, etc. It's just a question of whether the incremental safety of knowing about hazards outweighs the potential for intentional mischief.
Look at it this way; if you get false positives about hazards in front of you, the consequence is maybe you slow down unnecessarily. If you get a false negative, that's really no different from no data at all - your car should behave the same.
Chances are self-driving cars will have signed code and unique crypto ids stored in security enclaves. Every manufacturer will be able to see exactly which car signed the message. If someone was sending bad messages they can just ignore that id, or more likely disable the car and send the location to the police.
Heavy fog and rain would most likely be bad for Waymo's camera systems and its LiDAR systems, but Waymo cars also have radar, which would probably keep working pretty well. I'm not sure how well the car would perform if only radar was working well, though.
Are these cars using only visual data? I thought many used some kind of LIDAR as well, where that might cut through visibility issues depending on the wavelength used.
> autonomous driving tech is not to useful if it craps out in inclement weather.
This is the wrong way to think about it. An autonomous transportation service that works only in great weather is already a multibillion dollar opportunity. The revenue from providing services in places like California and Arizona can fund expansion into harder areas.
Nobody said that PCs were useless in the 1980s because many parts of the world didn't have reliable electricity. Nobody said web search was useless in the 1990s because most of the world didn't have Internet access. Likewise, you don't have to solve for driving in monsoon rains or Indian traffic before you can print money with autonomous vehicles.
I feel like this is a nice start, but Canada and more mountainous regions of the PNW will be the real proving ground. North Dakota, Minnesota and so on. At least in terms of generalized technical achievement, not in serving the most people in most situations. Winnipeg is a hellscape for driving in the best of times. Driving through the prairies in a snowstorm was one of the scariest experiences of my life, as well as the rockies with a thin layer of ice, intense grades, little to no road markings.
As a recognized tangent to that tangent, there's a big problem up here in the real North ( where basic necessities of life are ludicrously expensive. In part, at least possibly, due to transport costs. Would be amazing if self-driving vehicles could regularly deliver goods up there with no human risk on ice-roads. May be a pipe dream though.
I'm curious if we'll see the Detroit economy take off as a result of more self-driving car work in the area. It seems like Metro Detroit hasn't seen many of the tech gains that the coastal tech hubs have seen.
Here's my bet. Their software designed to drive conservatively and likewise it will be way too hair trigger with the brakes for use in the snow without bringing a lot of changes of pants.
That's assuming they can deal reasonably well with the reduced visibility and change in behavior of other drivers.
They'll probably also slide on a few right turns before they get the settings fully dialed in.
I get a chuckle out of these threads. Humans are _unbelievably_ bad at driving in the snow. If the only thing a self-driving car does in a whiteout snow storm is come to a halt, it will already be dramatically safer than human drivers. You can just hard-wire the car to not do stupid things. Taking human judgement out of the loop has many wonderful outcomes.
This is a bit aside from your main point, but stopping on a road in a whiteout snow storm is _incredibly_ dangerous due to the cars behind you being unable to see you until it is much too late to stop. I grew up driving two-lane highways in Alaska with frequent semi traffic, several times in blizzard conditions (to the point where the only way to follow the road was to drive on the flat part of the just-fallen snow). You could pull over to the side of the road, but you may not be able to get back on the road depending on your vehicle and the snow level (and in remote, cold areas this may be deadly itself).
You could say "just don't drive in blizzards", but on a long drive you may often end up in a blizzard when you started in clear weather. Frequently this happens in areas where there are no safe places to stop for miles.
I'm not an expert on self-driving car sensors, but it's not hard to imagine that they could be made much less weather-dependent than the human eye. Driving through blizzard conditions is incredibly disorienting and terrifying, so the potential of better-than-vision detection of roads and traffic is IMO the most promising way for self-driving cars to handle snowy conditions.
Potentially, a self-driving car (or even Google Maps nav and a regular driver) would not have this problem - it would simply not drive itself into a situation where it's both in heavy snow and a place that's not safe to stop. I've had a few times in California where I started a long drive in sun and warm weather and ended up stuck in a snow-storm mid-drive (hello Sierra passes)... it would have been nice if either my car or my map software warned me this was likely before we set off.
Autonomous cars would be coming from a variety of vendors each with its own algorithms. It's not even guaranteed that they would communicate with each other to arbitrate strategy under adverse conditions.
It's not as if human drivers do, either. The closest we've come is Waze, which would a trivial data set for an autonomous car to both utilize and supplement.
Self-driving cars feel like they are still in the early Alpha stage, while self-driving robots are already in production in Amazon's warehouses.
Why are companies going through the great expense of constructing full-size mock-ups if the domain problems haven't been solved on smaller mock-ups? Is there a fundamental difference between a 400 lb self-driving robot and a 2-ton self-driving robot - beyond the advertising novelty of being inside the car? Edit: If both robots are outside on a driving track, not in the Amazon warehouse.
And if "snow-vision" and other problems have been solved with smaller robots, or in other domains, why do news articles like this treat the topic as though it's entirely foreign and to be feared, instead of helping readers associate these concepts with existing concepts to drive acceptance?
Does an Amazon warehouse have thousands of pedestrians and high speed moving objects whizzing around them constantly, unknown regulations on their movement only visible via signs, and constantly evolving route types/conditions? Do Amazon robots need to figure all of these things out while moving at 80mph? Software for a robot to carry packages around a warehouse that was constructed for that purpose isn't even comparable to self driving cars.
The self driving robots have to have facilities specifically constructed and/or configured with the constraints of the robotic system in mind. No one makes a go anywhere, do anything self navigating robot.
The application of robotics in industry is still very much more akin to industrial automation than it is true autonomy. There's a ton you have to take into account to design a system that works well. In practice, this means that traditional automation is typically still the better solution than full autonomy.
I'm not sure I totally understand your point but a) a robot scooting around a flat concrete floor in a controlled environment is completely different from a car driving on real roads, and b) if they're ready to start full-size tests in these adverse conditions then they must think they're pretty darn close to having a viable product, given the possible consequences of failure. It's not like they haven't already done a massive amount of R&D to get to this point.
Again, I don't see your point. Before selling a self-driving car product intended to work reliably in X conditions, you have to test prototypes and eventually the production model "release candidate" in X conditions. I feel like you're trying too hard to find a cynical angle on this.
I’m confused. What kind of smaller mock-up are you imagining? I bet that Waymo considers lots of cases of driving solved (controlled environment parking lot) and we are seeing just another step up in terms of the scale of the problem.
How about interacting with pedestrians and cyclists? They could start with couriers, food delivery, animal control, city litter pick-up, janitorial services, etc. Of course they won't do that, because it's not "disrupting mobility," but merely eliminating some tedious, low-paying jobs.
If you're just working on computer vision and world mapping, why do you need something bigger than an RC car? I started with a sonar sensor and an arduino on an RC car in 2012, and had a working prototype, but never secured funding (and eventually became homeless).
Because an RC car with a sonar sensor and an arduino relates to a self driving car roughly the way a bicycle lock compares to a bank vault.
Orders of magnitude in complexity difference and a much more hostile environment to survive in.
I am not really comfortable writing this but seriously, if you have to ask this question I am actually happy you did not receive funding (but I am obviously not happy that you ended up homeless), that was not a prototype, it was not even a proof of concept. It was maybe a small step up from a toy. Sonar as a sensor may give you a little bit of input data but your arduino does not have the processing power to extract meaningful information about its environment from that data nor does that sensor give you sufficient information that you would use to pilot a couple of tons of steel through an environment inhabited by fragile humans.
OP sounds like that semi-autistic guy who asked Elon Musk if they could work together at the shareholder meeting. In his mind he's exaggerating his work/abilities while dragging down what Waymo and self-driving researchers are doing.
In the case of a real car, you aren't just reacting to the environment. The actors in the environment (cars, pedestrians, traffic light sensors) are reacting to you as well.
well that indoor robot has the advantage of fully mapped environment to operate within, low speeds, and people who know how to work around them.
going to be fun watching them deal with snow. though their approach has tended towards detail mapping. someone is going to figure out a break through in marking roads or perhaps they will get sensible and start with limited access highways first.
IR is attenuated in rain and fog, and I would presume snow as well. In many situations, there's no difference in resolution between IR sensors and human vision. [0]
Snow is increased visual noise. How do humans deal with it? We disregard it. How difficult is it to program computer vision to do the same?
Self-driving cars are at the intersection of numerous solved domains, yet integrating these remains strangely difficult for the largest companies in the world.
How can you so blindly oversimplify a problem and then act like the rest of the world is incompetent for not solving it already?
Snow is not just increased visual noise. Just off the top of my head, with no experience in self-driving cars and limited experience in computer vision:
Snow (on the ground and street and curbs) is a vastly lower-contrast visual environment where numerous traditionally distinct features of the landscape are covered by a blanket of white.
Snow is a more dangerous and difficult-to-stop-in environment, with totally different physical dynamics. Recognizing pedestrians (and their intentions), road blocks, debris, etc. is probably already a challenging problem, and with snow it's not only more difficult, it needs to be done from a much farther distance to give adequate stopping time.
Along with snow there are issues with recognizing icy patches and other especially low traction sections on the road and adjusting control inputs accordingly.
I can't tell if you've never driven in snow or never even attempted a basic computer vision or robotics task in your life, or both.
I don't think it's that. It's a fundamental misreading of the problem snow causes. The part where a few snowflakes go past the camera is relatively simple to deal with. But it's not why snowy conditions are difficult.
>Is there a fundamental difference between a 400 lb self-driving robot and a 2-ton self-driving robot
There's a difference in environment. How can you seriously think "Well if they have self-driving robots in controlled warehouse scenarios, what's the novelty of self-driving robots in unpredictable and often treacherous real-world public road scenarios?"
I think there are two separate issues that you raise that are essentially independent:
1. What evaluation metrics should we use to measure the advancement of self-driving technology?
2. How can we measure the public's level of acceptance for these technologies, and what impact does media coverage (positive or negative) have in that respect?
For example, I would not personally gauge that SDCs are in early alpha. However, I live in a suburb of Phoenix, where Uber, Waymo, and Cruise [0] are all conducting extensive testing; I usually observe between 15 and 30 distinct SDCs during my 20-minute commute. I expect this familiarity has (by design, I'm sure) made me more comfortable with SDCs than most. I also assume that the environment is somewhat simpler than other metro areas, as we have very little precipitation and a mostly grid-like street layout.
My judgement of these cars is that they approximate the behavior of an extremely stodgy, overly conservative, but ultimately very alert driver. For example, they respond well to being cut off, dealing with late light changes, or unpredictable pedestrian behavior, while dealing poorly with things which require decision-making, such as making a right on red, or preemptively changing lanes to avoid the frequent stops of a bus. In other words, they've been optimizing for safety, and not trip efficiency; I think that this is probably the right order of priorities at this stage of development.
Given this, I would be comfortable taking driverless trips around town, on surface streets; I would therefore classify the current state of development as early beta more so than early alpha. However, this is a subjective assessment, not based on any objective metric.
At the same time, I read the news too, and I think we are observing the clash of two competing agendas. On one hand, you have the tech companies painting an overly-rosy picture of development, probably to soften up the market to increase (eventual) penetration.
On the other hand, you have overly-pessimistic outlets preaching various types of doom (job loss, trolley problem, etc) in order to drive clicks through fear, or (cynically) to poison the market and buy time for later technical entries to catch up to the leaders.
Both of these tactics will have varying levels of receptivity with different groups, roughly corresponding (I expect) to the Rogers adoption curve [1]. Therefore, I expect we'll continue to see both optimistic and pessimistic coverage for quite some time.
[0]: In this order of development progress, in my opinion
I forget that truckers and delivery drivers see these not as angels coming to save them from the tyranny of driving, but rather see them as devils, coming to steal their jobs away. And so media outlets have to "wear kid gloves" when reporting these articles.
Same with the 'stock checking' bots at Walmart. They didn't develop that platform to check stock levels, they developed it to be a shop assistant. But they have to field test it and they can't afford to do so openly while a strike could still shut them down. Hence, "stock taking".
HN is a pretty critical place most of the time but sometimes it's valuable to take a step back. The challenges associated with building a car that can drive itself are pretty insanely difficult (even something as straightforward as "how do we get the training data?") and the fact that we're now at a point where there's a possibility of self-driving cars working in snowy conditions is pretty awesome.