Let's be honest, humans are pretty bad at driving in snow too.
Sadly, I feel like people won't accept self driving cars until they can drive in the snow as well as in dry, clear conditions, which won't happen for a long time, if ever.
Honestly it's a bit silly that self driving cars have to do better than humans to be trusted -- as long as they are doing as well as us, mathematically, we should want them.
Maybe the insurance companies, which are mathematically inclined, will end up fixing this by providing lower rates for self driving cars, even in snowy climates.
You want to live in a bubble, good for you. Give me freedom or give me death. I'd bet most of the people who are hard for these cars don't even own them, they just don't want to deal with taxi drivers anymore.
I pulled the CDC stats and it looks like eating and lack-of-exercise kill about 50x as many people as vehicle accidents.
Also generic falls are only 4K behind vehicle accidents. Somehow poisoning beat vehicle accidents by 4K, I'm not sure how, maybe counting prescription drug interactions and allergies?
It would appear the most dangerous daily activity is eating and not exercising. So your kitchen and living room couch are by far the most dangerous things in your life. Deep in the decimal points come poisoning, falling, and driving, all roughly the same odds. Someone else killing you is about half the odds of those three. Drowning is about a tenth the odds of driving although most adult drowning is probably outside the house and most people don't swim every day (although it is excellent exercise). On average you're about a fifth as likely to die of HIV than driving so the bedroom is probably safer than the drivers seat, on a very long term average, although HIV clusters extremely strongly into certain demographics, so bedrooms are basically safe for almost all of us and a death sentence for certain demographics.
In summary, there are some daily household activities that are safer than driving, such as having recreational sex, but your kitchen and living room couch are maybe a hundred times more likely to kill you than driving, and its about equal odds for falls, poisoning, and driving. The odds are about 49 in 50 that you're not going to die in a car accident.
(edited to add this is worrisome WRT death rates if people sit in self driving cars and stuff themselves with junk food while watching TV or something... the overall death rate might very well go up, even if vehicular deaths supernaturally dropped to zero)
People are already stuffing themselves with junk food while sitting in cars. Yes, even while driving - polls show that as many as half of drivers regularly eat while driving.
The problem is with humans it's fairly clear most of the time who made the error. It's either human 1 or human 2 99% of the time. With self driving cars it becomes software 1 or software 2... And at that point it's extremely difficult where to point the blame.
do you care to fill in the logic between your start and end? I don't really follow the leap. It seems the tools we use to determine fault now, frail though they may be, are equally sufficient no matter which bundles of electrical fields are controlling things.
Why would it be harder? Most likely there will be laws to disclose telemetry, video, etc. in the event of an accident. Most likely an audit trail of what the sensor inputs were and what the decision made was, etc.
More likely there will just be no-fault where both parties pay unless there is a spate of accidents that point to software/hardware issues.
I was recently wondering how the insurance industry will respond to self-driving cars. When self-driving vehicles are a small percentage of the market, it helps reduce their claims. When the entire grid is self-driving, the industry will be much smaller. We're probably at least 40 years away from that, but it still should happen someday.
Why do they need insurance? Poorly maintained cars will probably just refuse to go, so any flaws will be either deliberate tampering (which most insurances already don't cover, AFAIK) or the fault of the manufacturer.
That covers fleets, where maintenance can be strictly done according to the manufacturer's recommended schedule. Therefore, the only remaining entity that can be liable is the manufacturer.
Individually owned vehicles are not necessarily maintained to schedule. I would bet that a large percentage don't receive much attention at all, other than oil and filter, unless something goes wrong. Therefore, personal liability won't entirely go away.
For a self-driving car, programmed to follow traffic laws to the letter, an at-fault collision indicates a problem with the vehicle, either caused by a manufacturer defect (which includes software bugs) or by maintenance negligence by the owner. If an owner of a vehicle maintains it to the recommendations of the manufacturer, then any breakdowns also become manufacturer defects.
If a company owns a fleet of vehicles, and maintains them according to the manufacturer's recommendations, it is now impossible for them to be liable. A collision is always either the fault of the manufacturer or of another vehicle. Therefore, the owner of the vehicles need not have insurance. (This is not the current legal situation, but I expect that it would be in a few years.)
An individual person who owns a self-driving car would have exactly the same situation as above, provided they maintain the car to the recommended standard. In an attempt to deny a claim, the manufacturer would have the ability to argue that the vehicle was not maintained to standard, in which case the person would be liable for the results of that negligence.
No, but in general, failures of a machine that result in serious injury or fatality routinely result in lawsuits against the manufacturer. See GM ignition switches. Pharmaceuticals are probably the closest thing to an exception but drug interactions can result in lawsuits as well.
Or the system for determining fault could be flawed.
It would certainly be a legal problem for the manufacturer, but it doesn't necessarily point to a fault in the autonomous driving system.
I sort of expect that licensing for autonomous systems will in fact be fuzzy and require operators to cover liability, not require manufacturers to make flawless systems (whatever that would mean).
It'd seem the insurance rates would eventually be standardized to per car & driving technology; rather than taking into account any factors about the driver.
Let's be honest, humans are pretty bad at driving in snow too.
Sure, but most human drivers can call upon a wealth of experience both with driving and interacting with the physical world in general. They have an incredibly rich mental model, allowing for adaptations that would be risky or impossible to program into an autonomous driver. For example, when lane lines are missing or obscured, human drivers will often follow the car in front of them... unless they observe erratic behavior that might indicate the driver ahead has lost control. What's "erratic behavior" is a judgement I wouldn't expect near-future software to be particularly good at.
Of course, there are areas like reaction time and feedback-based control where autonomous drivers will always be vastly superior to human ones. Which is why it seems to me that the safest option in the medium term is likely to be (as it is today) an increasingly software-assisted human driver. This is the primary reason I'm down on the chances for wide adoption of fully (or mostly) autonomous drivers in the next 5-10 years: Human + computer will be a safer, more reliable alternative. The active participation of the human will fall off over time, but I think it will be a much flatter curve than many enthusiasts expect.
"Human + computer will be a safer, more reliable alternative."
Unfortunately, there's a cliff where if you aren't stimulating the human enough, they stop paying attention. They can't help it. Their neural architecture is fundamentally structured to do that and merely trying not to isn't going to cut it.
We can't have automated cars that 98% drive autonomously but beep when they need help, right now. This could work if the car can reliably predict when it is going to need help in 10+ seconds, but I'm not sure that's practical. (Which I phrase that way on purpose... maybe it is. I'm just not sure. Would certainly take a lot of work and some very conservative automated driving.)
To be able to know that it's doing something wrong or mis-reading things would require some deeper understanding of what it's doing (or at least a confidence score or similar). This is why it'd likely be important for s human to be overseeing it 100% of the time, to step in and swiftly make a manual override when the autonomy doesn't realize it's about to do something stupid.
A lot of people don't really oversee their car's driving 100% of the time today because they're fiddling with their phones or whatever. It's completely unrealistic to expect that anyone will actively monitor their environment if there's an autopilot that works reliably the vast majority of the time. Just won't/can't happen--and has been shown in research. Indeed, why would I even want a fully autonomous driving system if I have to drive the car anyway?
> Sadly, I feel like people won't accept self driving cars until they can drive in the snow as well as in dry, clear conditions, which won't happen for a long time, if ever.
Well yeah, I want to be able to drive my car year round. If it doesn't work in the snow it's useless for months for a large portion of the country. Even here in Seattle the rain is a big challenge.
I have a Mercedes Benz with adaptive cruise control, and was driving on the highway this weekend in a snowstorm. At one point, it basically gave up. It gave me an error message to the effect of "you're on your own now."
It is a challenge for a human, I feel for the algorithm folks trying to test the obscured vision.
How did the car alert you that you were on your own (I don't own a Mercedes)? I'm curious as to how it indicates to you a rather important message about your immediate driving situation.
I don't have a Mercedes but my Jeep Cherokee Trailhawk has this same functionality and I did a cross country trip through northern america for the holidays. On mine at least it says something like "Sensor obscured, adaptive cruise control/lane keeping disabled. Please wipe off sensors." as well as making an audible ding to alert me to the message in my dash display.
I don't own a Mercedes and I don't know how the car alerts you, but all that's really required in this situation is a ding and a message on the dash that says "Cruise control disabled" and then it turns off cruise control.
It makes the steering wheel shake or shudder, in a way that (to me at least) is meant to feel like hitting rumble strips at the side of the road. I think it probably dinged as well. It also turned off the cruise control, so I started losing speed. And there was an error message on the dashboard. So in all, I was well informed.
Woah, you drive with your cruise control on in bad weather conditions like that? I was always trained to never, ever use cruise control in adverse weather conditions. Without having your foot on the accelerator to be able to quickly react to a skid or whatever just seems insane...
Well the car would automatically take care of the skid. It's not cruise control like you're thinking of (it's radar based) and like any other modern MB it has advanced traction control. It even bowed out when it couldn't perform optimally. Sounds fine to me!
Simple cruise systems only monitored speed and would throttle the engine to maintain the target speed. When a loss of traction happens, the proper response is usually letting up on the throttle so traction can be regained. These old, simple systems would apply more throttle insuring a mere slip turns into an epic spinoff event.
Secondly, when driving in spotty traction conditions, it is good to use a lower numbered gear and a safe speed combined. That same loss of traction event can result in a very rapid spin of the tires due to the higher gears and torque combining to produce a rapid increase in tire spin rate when not constrained by the weight of the car and traction.
This can happen before a human can respond.
The higher gear puts the engine at a higher rpm and closer to it's natural unconstrained speed, which diminished both the event and it's impact considerably.
The human experiences momentary chatter as the grip of the tire varies, but does not experience runaway spin and the abrupt loss of control it brings.
I would think an automated car could be programmed with this all in mind and do rather well.
An awful lot of humans could use this kind of data programmed in as well. :)
I'm curious how this will perform in the real world since humans also cannot see the lane markings but, unlike the system described in the article, do not refer to stored or ad hoc environment mapping and instead operate on an entirely improvised "lanes by consensus".
No. Ad hoc environment mapping seems to imply the driverless cars will map the lanes have stored in their memory to the snowy road. Human drivers more or less invent new lanes by consensus.
Have they tested a driverless car "push it out of a snow-bank" mode yet? :)
For anyone who's lived in areas that have regular snow storms you know that eventually mutli-lane highways turn into single lane tracks of clear road. Usually these tracks are somewhere in the middle of the road and everyone just sticks to them because a semi or snow plow cleared it for us.
The method described is a really great way for car's to get stuck in snow drifts by trying to stay in lanes. Near term drivers will PROBABLY just take over in those situations but long term we'll probably need a combination of Lidar and Cameras to do this type of work.
The "lane lines optional" nature of snowstorms definitely is going to require a leap of logic I don't soon expect self-driving cars to understand.
Will the car be able to measure the depth of snow on the road? And find the most driveable part of the road? The high reflectivity of snow probably isn't a good ally for this use case.
Not to mention the composition or what might be below the snow.
Snow and ice are by far the hardest conditions for anyone driving. I have no doubt a computer can figure it out and be better eventually, until then it will be interesting.
Electronic stability control is required by regulation in several jurisdictions (US, Canada, EU now or soon). From that it can be inferred that computers are better at some aspects of it already.
I don't know what you think ESP is but it's purely a "damage control" type of driver aid and is only a layer of programming on top of ABS.
ESP just modulates brakes to keep you from going sideways. It's basically automated cutting brakes paired with traction control and ABS.
All ESP does is utilize the car's existing technology to better do what the driver wants. It pretty much only comes into affect when the driver cranks the wheel hard enough to use up all available traction. It keeps the vehicle in an understeer condition and uses all available traction to try and steer in the desired direction (by acting like cutting brakes).
The point of this is so that the "AWD means I don't actually need to know what I'm doing people" are prevented from sideways (oversteer) because understeer is a lot easier for a novice to recover from and if he/she hits a pole he/she hits it head on utilizing the part of the car that's good at crashing "softly"
The traction control part of ESP is somewhat irrelevant. If the used is trying to give it throttle while the ESP code is doing it's thing and the ESP code is telling the traction control code to override the user then the user should just pull the fuse because you do better doughnuts that way...
I understand what it is. It works because it can accurately estimate the available traction. A lot of the problems drivers have in slippery conditions boil down to not correctly estimating available traction.
> Not to mention the composition or what might be below the snow.
I don't feel like this is particularly harder than any of the other problems self-driving cars face. To achieve performance on par with human drivers simple need to do a couple of things (which are what I do when driving in snowy/icey conditions):
1) Categorize current/recent road appearance into groups.
2) Do tiny periodic break checks on each relevant group to determine the traction available in that group.
3) Trigger a new set break checks when conditions change (e.g. exterior temp changes, precipitation changes, traffic density changes or road-type changes such as exit ramp or bridge)
This road surface analysis combined with a bit of traffic analysis should also allow you to determine 'emergent lanes' in snowy conditions.
> Will the car be able to measure the depth of snow on the road? And find the most driveable part of the road? The high reflectivity of snow probably isn't a good ally for this use case.
Use a thermal camera to find the road? Or perhaps contrast, as the parts of the road with less slow should have roadway peek through. To determine how much snow you're dealing with, you detect wheel slip, which can assist in determining depth.
It's hard to suggest how this would be handled on a all-encompassing scale. I'm sure Google would like to simply track every car on the road, but between multiple brands of cars and software, and valid privacy concerns, this isn't a great solution. Also, you'd need to validate that data carefully. You wouldn't want false submissions to end up strongly recommend cars drive off a cliff.
So nationalize it? It's not like big brother isn't already tracking traffic trends. Why shouldn't the Dept of Transportation be in charge of directing automated traffic?
There's ups and downs to that. Companies like Google would argue that their systems are too integrated to their business to be nationalized. (Unless we just nationalize Google. It's enough of a monopoly now........)
The biggest issue would be that this is still a rapidly developing field, and the one thing government is not good at is rapidly evolving technology.
But fundamentally, we will need something controlled by an agency or government, or at least a nonprofit organization of some sort. A single corporation can't control our nation's transportation infrastructure.
Humans haven't been able to figure it out either (so many Winter accidents). At least the automated car will drive slowly and at a safe following distance.
Interestingly, driver aids have been implicated in causing more severe winter accidents, eg. this [1] 100-car pile-up in Sweden a few years back. Basically, people don't understand that the ESP lamp flashing means "slow down or you will probably crash". Drive an old, RWD, no electronics car and you'll be more aware of dangerous conditions. 100% non-overridable self-driving cars may someday fix this, but things will probably get worse before they get better. Humans being humans, they will think the car is overly cautious most of the time, and put it in manual because they're running 5 mins late.
The more interesting question about the automated car is "can it do a Scandinavian flick"? They should come with a dedicated button for that.
> driver aids have been implicated in causing more severe winter accidents, eg. this [1] 100-car pile-up in Sweden a few years back.
The linked article, AFAICT, doesn't implicate driver aids in the accident, at all. (It's a rather short article, and the video doesn't mention anything about driver aids either.) Am I missing something?
I couldn't find anything in English, but local media had discussions along these lines at the time, with incident responders blaming ESP systems. They had ambulances with flashing lights being hit by new cars coming to the scene of the accident, i.e. it didn't happen within just a few seconds; in this case it's apparent that drivers were oblivious to just how bad the conditions were.
I'm not saying ESP is a bad thing, mind you. The technical system works well and has reduced accidents overall, this is proven in several studies. There's been a few faulty ESP systems causing accidents, e.g. the Audi A3 and the Alfa Romeo Spider around '05 (fixed in recalls), but overall it's a success. But the more abstractions you put between the car and the driver, the more your run the risk of drivers ignoring, or not being aware of, bad conditions. Educating users about proper operation of any system is notoriously difficult.
It seems like the high reflectivity of snow would help in case of following the tracks of the previous car, where I'd guess the car's tracks are a little bit less reflective than the fresh untouched snow on the rest of the road?
You'd hope, but many area's "snow plowing" means the locals get together and hire someone with a road grating tractor to come and plow. AND if they have the extra budget toss some dirt on there.
:-) Obviously this doesn't apply to country or greater highways.
The majority of driving, probably. The majority of avoiding death and injury while driving, especially in snow, is to NOT follow the car in front of you.
As far as error rates, it's feels significant to me. I've driven maybe 500k miles in my lifetime and would be very surprised if I haven't seen more than 50 cars ahead of me spin or slide off of highways.
As a person that has had to learn to drive on the "wrong" side of the road, let me tell you that works really well, until it doesn't.
There is nothing more nerve wracking than arriving at a huge intersection with multiple turn lanes going onto bridges or into tunnels and being the first car sitting at the red.
Hopefully the car makers will be able to use data from the times when the driver needs to take over to build better image recognition of the ideal path in snowy roads.
It's strange to me how these conversations tend to center around failure modes that would be apparent to
a) anyone who has driven in that situation since it's not even remotely an outlier of an event
b) anyone who took their self driving car out into that situation more than a handful of times
we've already seen self driving cars leave their lanes in response to pedestrians or road hazards or just temporary detours around construction. Why would a snow bank not be avoidable as well?
As someone that grew up driving in the midwest I get a little nervous about autonomous driving being developed in California. It's a bit of a cliche, but CA drivers are notoriously bad drivers in winter conditions.
Of course the rational side of my brain knows there are brilliant people from all over the country (world?) working on autonomous driving for these companies, and they will have to do testing in all weather conditions across the whole country.
Perhaps it would help relax fears if companies like Google or Tesla opened (and publicised) an all-weather driving lab in someplace like northern Michigan.
Having spent multiple winters in Pittsburgh, it's not like a winter in the Rockies or the Tetons where as the GGGP describes, there are only single plowed lanes in multilane highways because it's snowing every day for two weeks straight. In Pittsburgh, the lanes get plowed fairly quickly and salted to death.
I suppose I was thinking more about the snow storms up near Lake Erie, which is only about 2 hours from Pittsburgh. Living in Erie, there was a few storms that led to I-90 having whiteout conditions and single lane traffic. Winters in Pittsburgh are fairly mild in comparison.
Having never lived near the Rockies, I don't know how it compares. They probably are worse though.
Tahoe is only a few hours away from Bay Area and virtually everyone* in bay area goes there at least a few times a year to ski. This year Tahoe got a lot of snow.
Yeah I mean Tahoe gets crazy snow, but there are some things that happen to the road at super cold temperatures mixed with varying levels of rain or snow.
I think these mixtures are more prevalent east snd north of CA.
It's not that usual for large projects to proceed a fair distance before seemingly obvious problems become apparent. Bridges that fail to meet in the center come to mind.
I remember being at a meeting talking about the implementation of a chipset where each of the three chips required another chip to be running first. Global conditions can be more obvious to an outsider than those involved in many details of creating X big plan.
Why would a snow bank not be avoidable as well?
The Google car in particular is intended to follow a pre-defined path derived from more-detailed-version of Google Maps with obstacle avoidance being just deviations from that pre-defined plan. It seems like the situation described by the op, the road-structure itself changing, would present significant obstacles to that strategy. Essentially, rather than solving the problem of guiding a vehicle down the road using visible spectrum image-processing, Google avoided the problem by various stands-in; lidar for obstacles and satellite maps for positioning. Clever but that sort of thing will have failures that might be surprising - and surprisingly obvious.
Are self-driving snow plows realistic, even if just on highways? Would such plows eventually be more affordable, be able to run almost 24 hours a day and mean that more road is cleared?
(It never snows where I live so plowing or driving on snow is an alien concept to me.)
They're probably realistic, but I'm not sure they'd make a lot of sense for most places right off the bat. They really only need to be out on the roads right after / during a snow storm, which isn't 24/7. For some places in the midwest, this might even be once a week max for a few hours. A human will be a lot cheaper until the cost comes way down.
They may make sense for areas that get a ton of snow, and have snow storms that last days, but I find it hard to believe the price will be cheaper than a human for a while.
"Ford said it had instead programmed the Lidar sensors to detect landmarks above the ground, such as buildings and road signs."
So, in order to work, it needs a mapped area. They skipped the actual problems with driving in snowy conditions. I guess I'll have to wait a while longer for an actual solution. At least the didn't go the full GPS route since I've been on roads that are marked 10 yards away from their actual location.
It better not be a one-time precomputed step since the roads changes constantly. I'm pretty sure I've never tried to precompute placement of items in a dynamic system.
Nonsense. Roads change very rarely compared to the number of trips taken on them. This precompute needs to be done just once per change, just like building an index that will be used for many searches. The road changes themselves almost always happen in less inclement weather.
This is pretty much exactly the same thing that humans do. I drive much better on a road with missing lane markings if I have driven on it before, and just like me, these systems will have degraded performance without the precompute.
On the other hand, I have been confused by fairly simple changes to roads such as a two way street being made one-way (which a decent satnav would have been updated to know about as soon as the change was gazettes) or roadworks closing off a road that i regularly use (again, autonomous system would have been updared based on gazetted road closures).
From my understanding, this is required of nearly all 'self-driving cars'. A Slate article from 2014 stated that Google's cars wouldn't know a stoplight existed, and would simply run it, if it wasn't in Google Maps. And that a Google engineer said the level of mapping required by their cars at the time would be 'impractical' to do at a nationwide level.
These cars require a lot more detail than traditional Google Maps data provides. And there's a huge concern about how they'd adapt to changes in the roadway which may have not made it to Google's servers yet.
Its one thing, and expected, for the car to look for signs, but if self-driving cars all need perfect maps which I didn't know was a necessity then I don't see them being practical.
My understanding is that it 'looks for signs', if it expects them to be there. Obviously, it has to look at the stoplight to see what color it is. But it won't look at arbitrary space for a stoplight it isn't told exists.
Source, by the way: slate.com/articles/technology/technology/2014/10/google_self_driving_car_it_may_never_actually_happen.html
FWIW, a Googler has done a TED talk about having the car recognize police officers holding signs and temporary things like that, more recently, so they may have made strides in solving this problem.
But at it's core, the Self-Driving Car Project seems way too reliant on Google's cloud servers for my taste. I'd love to have a self-driving car, but it needs to be self-driving, not Google-driving.
Give it ten years, and that traffic light problem may solve itself.
I would not surprise me if we would eventually see every traffic light broadcast its GPS location, cycle pattern, and phase, either using radio or by modulating their light source.
That still is risky (hackers could hack the logic in a traffic light), but less risky than relying on one central authority to know what he whole world looks like.
Also, I don't think it is a good idea, but chances are police officers will have ways to stop self-driving car without giving visual signs.
It would definitely be useful for traffic lights to be beacons, but I think you might underestimate the government time scale for renovating and modernizing our infrastructure on a national scale.
That's a clever solution. Many lights have moved over to arrays of colored LEDs. The bulbs themselves could transmit data, similar to LiFi, to cars. The LED arrays could also have a LiFi receiver.
To prevent hacking, the data transmitted by the traffic lights could be digitally signed using the public key of the jurisdiction. If 2 way communications is required, vehicles could sign their data with a vehicle key too.
Something like this would be a good way to securely allow lights to change their behavior when emergency vehicles or mass transit vehicles approach, broadcast road warnings to all vehicles, broadcast amber alerts. It could also receive reports from vehicles on average speed in road segments, reports of icy conditions (from vehicle traction feedback), reports of potholes or other maintenance issues, etc.
Consider also: crossing guards, police pulling you or someone else over, emergency vehicles trying to get through (possibly in traffic necessitating that multiple cars move around to make way), drunk driver swerving into your lane, construction lanes which route you onto the shoulder, stoplight failure (power outage or otherwise) with and without someone directing traffic, city parking garages where a valet parks your car in a tight grid, events where you park in a field / on the grass (somewhere that would otherwise be "out of bounds"), camping, some of the things schools do inside parking lots...
I just see far too many weird edge cases for car without a steering wheel to be feasible in the next few decades. Automated driver assist, sure. Taking over almost all the time, sure. But the dream of having it take you home drunk / take your kids to school without you / let you sleep on the way to work is pretty far off IMO.
I agree, and most people don't understand the extent of parking being a huge issue. Driveways for houses, all of it.
Though there are some edge cases companies like Google might be able to do in the shorter term. An Uber-like service only needs to show you the closest it is able to pick you up and drop you off, and can largely ignore the parking nightmare entirely.
You'd be surprised. It's just not an issue. Google has already done it for Mountain View. Yes, any additional city will need to have an equivalent map built for it. Ok. Great! So that's what they'll do. It's a solved problem. Throw some time and money at it and voila.
In the aforementioned source, a Googler said that's impractical. A lot of things could be solved with an infinite amount of money, that we recognize is not solved because we don't have the money to do it.
Does that problem require better mapping or better reporting and data distribution by the relevant authorities that infrastructure has been installed/modified/removed?
This isn't a realistic consideration in any short-term future. That assumes 100% of cars on the road are self-driving cars which are talking to each other. That is decades away at the absolute best speed scenario. But self-driving cars won't make it to that point if they can't cope with the world we live in now.
Hopefully, you'll just cross anywhere, and they'll just stop for you. Maybe we can figure some way to indicate you want to cross, so that they can slow down smoothly - as posted above, Google software can already read hand signs by cyclists.
Standing facing the kerb and looking at the traffic is a pretty good way of indicating you want to cross. Not sure how it would handle the other factor though: eye contact.
> A Slate article from 2014 stated that Google's cars wouldn't know a stoplight existed, and would simply run it, if it wasn't in Google Maps
That muse be old. In their TED talk from last year, they had it spotting and reacting to cyclist hand signals, school bus mechanical stop signs, and even a police officer directing traffic at an intersection.
> In fact, when I first wrote about the car for MIT Technology Review, Google admitted to me that the process it currently uses to make the maps are too inefficient to work in the country as a whole.
but his previous article he links to says
> Maps have so far been prepared for only a few thousand miles of roadway, but achieving Google’s vision will require maintaining a constantly updating map of the nation’s millions of miles of roads and driveways. Urmson says Google’s researchers “don’t see any particular roadblocks” to accomplishing that.
(and, incidentally, you quote the word "impractical", which appears in neither quote :)
I guess I did translate 'inefficient' to 'impractical' in my head. I am not certain I see a large difference though. There are arguably many roads where getting a constantly updating map is relatively doable, and many where it is not. Even if the cars themselves are able to update the map, how will they handle the trailblazing through an inaccurately mapped area?
I also hesitate to trust the TED talk, to be honest. Even some of Google's earliest videos of their self-driving car project seemed to portray a level of capability beyond what their cars could actually do without a lot of hand-holding.
Every time a Google / Tesla / Ford / GM / Toyota self-driving car passes down a road in normal conditions, it will be mapping the surroundings in fairly intricate detail.
Based on GPS + Lidar + other sensors, the roadway and surroundings will be mapped to within centimeters -- which should make navigating during inclement conditions fairly straight-forward.
These systems might fail on snowy roads that have never been traversed previously, but what percentage of driving is performed on roads with hundreds/thousands of cars passing every day? 99.9%? 99.99%?
The comparison to make is not with yourself, it's with the worst drivers that are already legally allowed to drive. Getting them replaced with a superior autonomous system makes you safer regardless of how your skill level compares to the system.
(The only crash I've been in involved someone else losing control of their vehicle and crossing the center line on a major highway, which probably informs this perspective a little bit)
I think you are right about that being a possibility, but my personal judgment is that a lot of bad driving is rooted in poor decision making and concentrated, so I don't think an overall decrease in safety is a likely outcome.
% isn't really a meaningful way to express an accident rate. You probably want to use some measure like accident rate per vehicle mile. I don't know the accident rate but the fatality rate in the US is about 1 death per 100 million miles.
On a similar note: We have our highway fatality posted every day on those over-the-freeway signs that I read each morning. It averages almost exactly one per day, statewide.
But this year, here we are non day 11 and only 3 fatalities so far! I'm hopeful we can keep this rate through the spring.
Until that number is 100, a car without controls will fail catastrophically at some point and can't be fully autonomous, or will need some huge disclaimers. It's an extreme long-tail problem.
I agree there will be a lot of failures in self driving cars, especially in the short term - but these should be in the nature of graceful degradation.
In the long term, 4 or 5 nines will probably better than most drivers.
I think he is referring to incorrect map data, which would be much more dangerous than missing map data.
Leaving aside malicious map data for now, high-speed roadways would tend to be well mapped, and low speed roadways should be discoverable in real time.
What could be very dangerous is bad map data in addition to a construction zone or road washout. I do think that places like construction zones will be the first to have V2I (vehicle to infrastructure) installed - to be extra sure that vehicles are aware of unusual conditions.
I'm not sure why this is a problem. If a self-driving car encounters any combination of conditions / poorly-mapped roads where it cannot drive safely, it can:
1) Pull over / turn around
2) Pass full control to one of the passengers
3) Find a longer/slower route with sufficiently good conditions/mapping
It (potentially) solves one challenging problem - how to navigate when the roads are not clear or clearly marked. Unfortunately it's a fairly limited solution - "Falling rain and snow can interfere with LIDAR and cameras" (http://www.wired.com/2016/01/the-clever-way-fords-self-drivi...)
There are actually a lot of places that snow in California, you just have to get up into the mountains. There's a lot of snow within just a couple hours of Silicon Valley.
It's not that hard. You use LIDAR to generate a topographic map of the terrain ahead. We were doing that 10 years ago for the DARPA Grand Challenge. For off-road driving, this is essential. A high-mounted LIDAR lets you do this. It's not just about sensing obstacles. You profile the height of the terrain ahead, cell by cell.
There's a range limit on terrain profiling, though, imposed by parallax and the height of the scanner above the road. The higher, the better, of course. For a scanner 2m above the road, you can get good profiles out to 20m or so. This is enough to let you drive about 35- 40mph, which is enough for most heavy snow situations.
How much impact do rain and snow have on the LIDAR? That seems to come up all the time, and I've never seen a real definitive dismissal of the concern.
What you need for that is a LIDAR with "first and last". You want both the first return and the last return. Dust, rain, snow, and fog will produce noisy early returns, and solid terrain will produce the last return. You can also range-gate, ignoring returns out to a certain distance, and move the range gate in and out until you sense something solid.
LIDAR can be used to see through fog, snow, and dust to a limited extent.[1][2][3] Better than human vision or cameras.
Given all car makers are investing in driverless cars- why don't they install sensors or guide-wires into roads to mark lanes, etc. the next time they are repaving/repairing roads? won't that be much safer than having to rely on cameras/radars .
This has been proposed before, and the short answer is that it is too expensive to install these things into all of the roads.
Roads are typically resurfaced only once every 15-25 years, and in many municipalities, shrinking budgets mean that some roads are just outright being closed rather than being maintained.
More like 1-10 years. And the snow plover destroy the surface of the roads, so you can't mount some device on the road surface or just beneath the surface. Even the sticks net to the road get destroyed from time to time by trucks or snow plovers.
Give me a driverless car that only works on perfect sunny days, and the rest of the time requires me to drive it manually. That's still 10 x better than the car I have right now.
Firstly I can see any of these being fairly easily solvable. We're still in the early days in this tech. It will likely take decades to really nail it.
Secondly, even if half are solved, that's half of things I would have had to do anyway (well scraping ice of windshield instead of sensors).
Thirdly I do recognise there are issues e.g. "Ford said it had instead programmed the Lidar sensors to detect landmarks above the ground, such as buildings and road signs." This would still be problematic in heavy snow and away from distinct fixed landmarks.
But can we just appreciate the advancements that we are seeing year on year!
I think you have an autopilot-type system. This is driverless. That means nobody around. Nobody to do any of those things. So it is an all-or-nothing game.
The ideal with self-driving cars is that they're basically a fleet of taxis which are either constantly moving or parked in some central garage.
Airplanes deal with far worse sensor icing conditions every day. It's a problem that needs to be considered, but not a terribly complicated one. Cars have had heated windows for decades.
I suspect there is an issue with update frequency or spatial resolution, but it seems like it could be a good fit for the sort of best-effort approach described in OP.
"But these sensors do not work well in snowy conditions, and the car's onboard cameras cannot see road markings obscured by snow. Ford said it had instead programmed the Lidar sensors to detect landmarks above the ground, such as buildings and road signs."
Countryside usually there is a lot more snow on the street and especially next to the street. Detecting "landmarks" there will be hard.
Also if you combine LIDAR with stereo cameras, you will find out like Mercedes dod in the 1980s that cameras have problems with low stand sun (light rays), common in winter season (evening). So winter conditions will be harder to solve.
141 comments
[ 3.2 ms ] story [ 193 ms ] threadSadly, I feel like people won't accept self driving cars until they can drive in the snow as well as in dry, clear conditions, which won't happen for a long time, if ever.
Honestly it's a bit silly that self driving cars have to do better than humans to be trusted -- as long as they are doing as well as us, mathematically, we should want them.
Maybe the insurance companies, which are mathematically inclined, will end up fixing this by providing lower rates for self driving cars, even in snowy climates.
You can still walk.
They don't allow your rocket car on public roads today, so obviously there's some level of bubble you're willing to accept.
Also generic falls are only 4K behind vehicle accidents. Somehow poisoning beat vehicle accidents by 4K, I'm not sure how, maybe counting prescription drug interactions and allergies?
http://www.worldlifeexpectancy.com/usa-cause-of-death-by-age...
It would appear the most dangerous daily activity is eating and not exercising. So your kitchen and living room couch are by far the most dangerous things in your life. Deep in the decimal points come poisoning, falling, and driving, all roughly the same odds. Someone else killing you is about half the odds of those three. Drowning is about a tenth the odds of driving although most adult drowning is probably outside the house and most people don't swim every day (although it is excellent exercise). On average you're about a fifth as likely to die of HIV than driving so the bedroom is probably safer than the drivers seat, on a very long term average, although HIV clusters extremely strongly into certain demographics, so bedrooms are basically safe for almost all of us and a death sentence for certain demographics.
In summary, there are some daily household activities that are safer than driving, such as having recreational sex, but your kitchen and living room couch are maybe a hundred times more likely to kill you than driving, and its about equal odds for falls, poisoning, and driving. The odds are about 49 in 50 that you're not going to die in a car accident.
(edited to add this is worrisome WRT death rates if people sit in self driving cars and stuff themselves with junk food while watching TV or something... the overall death rate might very well go up, even if vehicular deaths supernaturally dropped to zero)
More likely there will just be no-fault where both parties pay unless there is a spate of accidents that point to software/hardware issues.
https://usinsuranceagents.com/answers/3251/what-is-self-insu...
Leasing capacity with little delay will only work in regions with dense population.
Individually owned vehicles are not necessarily maintained to schedule. I would bet that a large percentage don't receive much attention at all, other than oil and filter, unless something goes wrong. Therefore, personal liability won't entirely go away.
If a company owns a fleet of vehicles, and maintains them according to the manufacturer's recommendations, it is now impossible for them to be liable. A collision is always either the fault of the manufacturer or of another vehicle. Therefore, the owner of the vehicles need not have insurance. (This is not the current legal situation, but I expect that it would be in a few years.)
An individual person who owns a self-driving car would have exactly the same situation as above, provided they maintain the car to the recommended standard. In an attempt to deny a claim, the manufacturer would have the ability to argue that the vehicle was not maintained to standard, in which case the person would be liable for the results of that negligence.
It would certainly be a legal problem for the manufacturer, but it doesn't necessarily point to a fault in the autonomous driving system.
I sort of expect that licensing for autonomous systems will in fact be fuzzy and require operators to cover liability, not require manufacturers to make flawless systems (whatever that would mean).
That seems way too pessimistic.
Sure, but most human drivers can call upon a wealth of experience both with driving and interacting with the physical world in general. They have an incredibly rich mental model, allowing for adaptations that would be risky or impossible to program into an autonomous driver. For example, when lane lines are missing or obscured, human drivers will often follow the car in front of them... unless they observe erratic behavior that might indicate the driver ahead has lost control. What's "erratic behavior" is a judgement I wouldn't expect near-future software to be particularly good at.
Of course, there are areas like reaction time and feedback-based control where autonomous drivers will always be vastly superior to human ones. Which is why it seems to me that the safest option in the medium term is likely to be (as it is today) an increasingly software-assisted human driver. This is the primary reason I'm down on the chances for wide adoption of fully (or mostly) autonomous drivers in the next 5-10 years: Human + computer will be a safer, more reliable alternative. The active participation of the human will fall off over time, but I think it will be a much flatter curve than many enthusiasts expect.
Unfortunately, there's a cliff where if you aren't stimulating the human enough, they stop paying attention. They can't help it. Their neural architecture is fundamentally structured to do that and merely trying not to isn't going to cut it.
We can't have automated cars that 98% drive autonomously but beep when they need help, right now. This could work if the car can reliably predict when it is going to need help in 10+ seconds, but I'm not sure that's practical. (Which I phrase that way on purpose... maybe it is. I'm just not sure. Would certainly take a lot of work and some very conservative automated driving.)
> a judgement I wouldn't expect near-future software to be particularly good at.
Its going to take a fleet of a ten thousand self driving cars a week to have driving experience equal any person alive.
Thats why Google / et al are putting millions of miles of data into these logic engines before they even attempt to sell them.
Well yeah, I want to be able to drive my car year round. If it doesn't work in the snow it's useless for months for a large portion of the country. Even here in Seattle the rain is a big challenge.
It is a challenge for a human, I feel for the algorithm folks trying to test the obscured vision.
Secondly, when driving in spotty traction conditions, it is good to use a lower numbered gear and a safe speed combined. That same loss of traction event can result in a very rapid spin of the tires due to the higher gears and torque combining to produce a rapid increase in tire spin rate when not constrained by the weight of the car and traction.
This can happen before a human can respond.
The higher gear puts the engine at a higher rpm and closer to it's natural unconstrained speed, which diminished both the event and it's impact considerably.
The human experiences momentary chatter as the grip of the tire varies, but does not experience runaway spin and the abrupt loss of control it brings.
I would think an automated car could be programmed with this all in mind and do rather well.
An awful lot of humans could use this kind of data programmed in as well. :)
Have they tested a driverless car "push it out of a snow-bank" mode yet? :)
The method described is a really great way for car's to get stuck in snow drifts by trying to stay in lanes. Near term drivers will PROBABLY just take over in those situations but long term we'll probably need a combination of Lidar and Cameras to do this type of work.
Will the car be able to measure the depth of snow on the road? And find the most driveable part of the road? The high reflectivity of snow probably isn't a good ally for this use case.
Snow and ice are by far the hardest conditions for anyone driving. I have no doubt a computer can figure it out and be better eventually, until then it will be interesting.
ESP just modulates brakes to keep you from going sideways. It's basically automated cutting brakes paired with traction control and ABS.
All ESP does is utilize the car's existing technology to better do what the driver wants. It pretty much only comes into affect when the driver cranks the wheel hard enough to use up all available traction. It keeps the vehicle in an understeer condition and uses all available traction to try and steer in the desired direction (by acting like cutting brakes).
The point of this is so that the "AWD means I don't actually need to know what I'm doing people" are prevented from sideways (oversteer) because understeer is a lot easier for a novice to recover from and if he/she hits a pole he/she hits it head on utilizing the part of the car that's good at crashing "softly"
The traction control part of ESP is somewhat irrelevant. If the used is trying to give it throttle while the ESP code is doing it's thing and the ESP code is telling the traction control code to override the user then the user should just pull the fuse because you do better doughnuts that way...
I don't feel like this is particularly harder than any of the other problems self-driving cars face. To achieve performance on par with human drivers simple need to do a couple of things (which are what I do when driving in snowy/icey conditions):
1) Categorize current/recent road appearance into groups.
2) Do tiny periodic break checks on each relevant group to determine the traction available in that group.
3) Trigger a new set break checks when conditions change (e.g. exterior temp changes, precipitation changes, traffic density changes or road-type changes such as exit ramp or bridge)
This road surface analysis combined with a bit of traffic analysis should also allow you to determine 'emergent lanes' in snowy conditions.
Use a thermal camera to find the road? Or perhaps contrast, as the parts of the road with less slow should have roadway peek through. To determine how much snow you're dealing with, you detect wheel slip, which can assist in determining depth.
The biggest issue would be that this is still a rapidly developing field, and the one thing government is not good at is rapidly evolving technology.
But fundamentally, we will need something controlled by an agency or government, or at least a nonprofit organization of some sort. A single corporation can't control our nation's transportation infrastructure.
The more interesting question about the automated car is "can it do a Scandinavian flick"? They should come with a dedicated button for that.
[1] http://www.bbc.com/news/world-europe-21035634
The linked article, AFAICT, doesn't implicate driver aids in the accident, at all. (It's a rather short article, and the video doesn't mention anything about driver aids either.) Am I missing something?
I'm not saying ESP is a bad thing, mind you. The technical system works well and has reduced accidents overall, this is proven in several studies. There's been a few faulty ESP systems causing accidents, e.g. the Audi A3 and the Alfa Romeo Spider around '05 (fixed in recalls), but overall it's a success. But the more abstractions you put between the car and the driver, the more your run the risk of drivers ignoring, or not being aware of, bad conditions. Educating users about proper operation of any system is notoriously difficult.
I'm curious how they'll handle whiteout conditions, however.
:-) Obviously this doesn't apply to country or greater highways.
Edited to soften the first sentence, a quick search didn't find any active users, just research.
Just follow the car in front of you. That's what the majority of driving is anyway.
As far as error rates, it's feels significant to me. I've driven maybe 500k miles in my lifetime and would be very surprised if I haven't seen more than 50 cars ahead of me spin or slide off of highways.
There is nothing more nerve wracking than arriving at a huge intersection with multiple turn lanes going onto bridges or into tunnels and being the first car sitting at the red.
a) anyone who has driven in that situation since it's not even remotely an outlier of an event
b) anyone who took their self driving car out into that situation more than a handful of times
we've already seen self driving cars leave their lanes in response to pedestrians or road hazards or just temporary detours around construction. Why would a snow bank not be avoidable as well?
Of course the rational side of my brain knows there are brilliant people from all over the country (world?) working on autonomous driving for these companies, and they will have to do testing in all weather conditions across the whole country.
Perhaps it would help relax fears if companies like Google or Tesla opened (and publicised) an all-weather driving lab in someplace like northern Michigan.
[] - http://www.uberatc.com/ (which seems like a unofficial site but I can't find anything else)
Having never lived near the Rockies, I don't know how it compares. They probably are worse though.
*everyone = Google/Apple/FB employees
I think these mixtures are more prevalent east snd north of CA.
I remember being at a meeting talking about the implementation of a chipset where each of the three chips required another chip to be running first. Global conditions can be more obvious to an outsider than those involved in many details of creating X big plan.
Why would a snow bank not be avoidable as well?
The Google car in particular is intended to follow a pre-defined path derived from more-detailed-version of Google Maps with obstacle avoidance being just deviations from that pre-defined plan. It seems like the situation described by the op, the road-structure itself changing, would present significant obstacles to that strategy. Essentially, rather than solving the problem of guiding a vehicle down the road using visible spectrum image-processing, Google avoided the problem by various stands-in; lidar for obstacles and satellite maps for positioning. Clever but that sort of thing will have failures that might be surprising - and surprisingly obvious.
(It never snows where I live so plowing or driving on snow is an alien concept to me.)
They may make sense for areas that get a ton of snow, and have snow storms that last days, but I find it hard to believe the price will be cheaper than a human for a while.
So, in order to work, it needs a mapped area. They skipped the actual problems with driving in snowy conditions. I guess I'll have to wait a while longer for an actual solution. At least the didn't go the full GPS route since I've been on roads that are marked 10 yards away from their actual location.
This is pretty much exactly the same thing that humans do. I drive much better on a road with missing lane markings if I have driven on it before, and just like me, these systems will have degraded performance without the precompute.
These cars require a lot more detail than traditional Google Maps data provides. And there's a huge concern about how they'd adapt to changes in the roadway which may have not made it to Google's servers yet.
Source, by the way: slate.com/articles/technology/technology/2014/10/google_self_driving_car_it_may_never_actually_happen.html
FWIW, a Googler has done a TED talk about having the car recognize police officers holding signs and temporary things like that, more recently, so they may have made strides in solving this problem.
But at it's core, the Self-Driving Car Project seems way too reliant on Google's cloud servers for my taste. I'd love to have a self-driving car, but it needs to be self-driving, not Google-driving.
I would not surprise me if we would eventually see every traffic light broadcast its GPS location, cycle pattern, and phase, either using radio or by modulating their light source.
That still is risky (hackers could hack the logic in a traffic light), but less risky than relying on one central authority to know what he whole world looks like.
Also, I don't think it is a good idea, but chances are police officers will have ways to stop self-driving car without giving visual signs.
To prevent hacking, the data transmitted by the traffic lights could be digitally signed using the public key of the jurisdiction. If 2 way communications is required, vehicles could sign their data with a vehicle key too.
Something like this would be a good way to securely allow lights to change their behavior when emergency vehicles or mass transit vehicles approach, broadcast road warnings to all vehicles, broadcast amber alerts. It could also receive reports from vehicles on average speed in road segments, reports of icy conditions (from vehicle traction feedback), reports of potholes or other maintenance issues, etc.
I just see far too many weird edge cases for car without a steering wheel to be feasible in the next few decades. Automated driver assist, sure. Taking over almost all the time, sure. But the dream of having it take you home drunk / take your kids to school without you / let you sleep on the way to work is pretty far off IMO.
Though there are some edge cases companies like Google might be able to do in the shorter term. An Uber-like service only needs to show you the closest it is able to pick you up and drop you off, and can largely ignore the parking nightmare entirely.
That muse be old. In their TED talk from last year, they had it spotting and reacting to cyclist hand signals, school bus mechanical stop signs, and even a police officer directing traffic at an intersection.
https://www.youtube.com/watch?v=tiwVMrTLUWg
edit: I see you mention this below.
Weird article, though. The author says:
> In fact, when I first wrote about the car for MIT Technology Review, Google admitted to me that the process it currently uses to make the maps are too inefficient to work in the country as a whole.
but his previous article he links to says
> Maps have so far been prepared for only a few thousand miles of roadway, but achieving Google’s vision will require maintaining a constantly updating map of the nation’s millions of miles of roads and driveways. Urmson says Google’s researchers “don’t see any particular roadblocks” to accomplishing that.
(and, incidentally, you quote the word "impractical", which appears in neither quote :)
I also hesitate to trust the TED talk, to be honest. Even some of Google's earliest videos of their self-driving car project seemed to portray a level of capability beyond what their cars could actually do without a lot of hand-holding.
Based on GPS + Lidar + other sensors, the roadway and surroundings will be mapped to within centimeters -- which should make navigating during inclement conditions fairly straight-forward.
These systems might fail on snowy roads that have never been traversed previously, but what percentage of driving is performed on roads with hundreds/thousands of cars passing every day? 99.9%? 99.99%?
(The only crash I've been in involved someone else losing control of their vehicle and crossing the center line on a major highway, which probably informs this perspective a little bit)
But this year, here we are non day 11 and only 3 fatalities so far! I'm hopeful we can keep this rate through the spring.
In the long term, 4 or 5 nines will probably better than most drivers.
Leaving aside malicious map data for now, high-speed roadways would tend to be well mapped, and low speed roadways should be discoverable in real time.
What could be very dangerous is bad map data in addition to a construction zone or road washout. I do think that places like construction zones will be the first to have V2I (vehicle to infrastructure) installed - to be extra sure that vehicles are aware of unusual conditions.
1) Pull over / turn around
2) Pass full control to one of the passengers
3) Find a longer/slower route with sufficiently good conditions/mapping
There's a range limit on terrain profiling, though, imposed by parallax and the height of the scanner above the road. The higher, the better, of course. For a scanner 2m above the road, you can get good profiles out to 20m or so. This is enough to let you drive about 35- 40mph, which is enough for most heavy snow situations.
LIDAR can be used to see through fog, snow, and dust to a limited extent.[1][2][3] Better than human vision or cameras.
[1] http://viodi.com/2015/02/17/better-eyesight-than-humans-the-... [2] http://mil-embedded.com/articles/case-helicopter-pilots-clea... [3] http://www.laseroptronix.se/gated/Rangegatedcameras.pdf
Roads are typically resurfaced only once every 15-25 years, and in many municipalities, shrinking budgets mean that some roads are just outright being closed rather than being maintained.
Can it hook up its own jumper cables?
Can it scrape the ice off all those sensors?
Can it scrape the ice off the license plate, windows, lights and everything else needed to not get a ticket from an irate cop?
Can it stop to give a lift to a stranded person on christmas eve?
It can certainly navigate and move in snowy conditions, but I can't see it actually driving in them.
Firstly I can see any of these being fairly easily solvable. We're still in the early days in this tech. It will likely take decades to really nail it.
Secondly, even if half are solved, that's half of things I would have had to do anyway (well scraping ice of windshield instead of sensors).
Thirdly I do recognise there are issues e.g. "Ford said it had instead programmed the Lidar sensors to detect landmarks above the ground, such as buildings and road signs." This would still be problematic in heavy snow and away from distinct fixed landmarks.
But can we just appreciate the advancements that we are seeing year on year!
Airplanes deal with far worse sensor icing conditions every day. It's a problem that needs to be considered, but not a terribly complicated one. Cars have had heated windows for decades.
I suspect there is an issue with update frequency or spatial resolution, but it seems like it could be a good fit for the sort of best-effort approach described in OP.
Countryside usually there is a lot more snow on the street and especially next to the street. Detecting "landmarks" there will be hard.
Also if you combine LIDAR with stereo cameras, you will find out like Mercedes dod in the 1980s that cameras have problems with low stand sun (light rays), common in winter season (evening). So winter conditions will be harder to solve.