Self-Driving Cars Can Handle Neither Rain nor Sleet nor Snow
bloomberg.com
bloomberg.com
LIDAR units that return "first and last" returns are helpful in dealing with rain and fog. The first return (nearest thing seen) will be noisy in rain, but last (most distant thing in that direction) should be stable if it's from a solid object. "First and last" is used with aerial LIDAR surveying; "first" gives you the treetops, "last" gives you the ground.
Range gated imagers for seeing through fog and dust have been available for over a decade. These are active devices which allow filtering out anything that's outside a narrow image range gate. You can adjust the gate and get images layer by layer. This technology is used by the military and seems to be somewhat restricted.[1][2][3] In addition to fog, it can show what's behind a camouflage net, objects concealed by brush, and such.
The submillimeter radar people are slowly making progress, too.[4] Automotive radars are currently running at 77GHz, which is low resolution compared to LIDAR. The resolution gets better with frequency. There are now systems in the 300GHz range.
So the sensors needed are available now, in expensive forms, and can get cheaper.
[1] http://www.dvsmil.com/PDF/Imaging_Lidar_Seeing_Through_Cloud... [2] https://www.youtube.com/watch?v=hOTo-6jXOYo [3] http://www.sensorsinc.com/applications/military/laser-range-... [4] http://consortis.eu/
Here's a video of what's claimed to be one of the higher resolution radars.[1] About one degree. It can see cars.
Many new startups in this area.[2] Most are operating in the 77GHz range, which is OK for sensing other cars but marginal for smaller targets. LIDAR-type resolution is still a ways off.
[1] https://www.youtube.com/watch?v=r0s5P1SQ34M [2] https://www.mwrf.com/systems/startups-trying-revamp-automoti...
nononono! Even assuming you can create sufficient map data as soon as there is any kind of roadworks - or even only the winter cracking the asphahlt - your data will be useless. They don't even need to change the layout of the road. I'll never understand companies that focus so much on maps as priamry source of data. Every day, massive - temporary (construction site) or permanent - changes are occurring to the road network. Also, the centimeter-precise location of the vehicle doesn't matter that much. The actual situation on the road matters much more: You may need to drive a meter more to the left because of a parked truck, a vehicle may obscure signs, a tree branch hangs down into the road, some car is parked too close to intersection, etc.
If I then drive the road and there's been some changes (traffic cones, lane closures, heavier traffic than normal, dark so I need to watch out for animals more closely....) I adjust my speed for the circumstances
I don't see why gathering surface information would be any different. Gather info, compare to previous fleet-gathered info, adjust driving to compensate. Seems reasonable to me (as ONE source of info for decision-making). And as fleet size increases, info gets gathered more quickly - look at how quickly google maps can inform you of traffic slowdowns in busy urban areas.
Maybe they don't really care about the product. Maybe the execs only care about making money, so they'll find funding for anything - whether it will pan out or not.
Self-driving cars are hot and for now there's a lot of money available. I think things will cool down when the hard part starts - like when they really do have to figure out how to make them drive in less than ideal conditions.
While not sufficient, this does play to one of the fundamental advantages computers have over people: you can copy information from one to the other.
The highway value is when snow covers lane markers. The GPS will still work, but it's only so accurate without those lane markers etc.
Then a single dead pedestrian changes all the news articles and the opinions change.
We are a very, very long way away from self driving taxi cabs. What a total PR scam that was - but it did help valuations.
How many deaths per year would be acceptable? Like you mention, in the US, it seems that zero deaths per year, similar to commercial aviation, would be required. It is a bit of a shame that a tech that could save a million lives a year globally won't be deployed because it can't be made perfectly safe. On the other hand, I'm sure people in India or China will be happy to build and deploy them with much higher failure rates. Maybe we can import those to the US someday.
This is a bit of a generous assumption. It very well could turn out that self-driving cars are worse, or no better, than human drivers.
Not to mention that the failure modes of self-driving cars are likely to be very different from failure modes of humans, making it more difficult for other road users to predict what they would do in any given situation.
The only question is can we make them work in situations that are hard.
In other words, "it works until it doesn't" is tautologically true, but useless precisely for that reason.
On the other hand the self-driving cars are just getting started. It's entirely reasonable to suppose that they may surpass the average human in the next X years. And they won't get old and confused and eventually replaced with new cars that have to learn from scratch -- they should get more or less monotonically better. I think the optimism is warranted even if the timeline and technology is uncertain.
This is trivially true. I mean, self driving cars that are worse than human drivers already exist. Isn’t the point asking when they will be made widely available? I assume that is only possible when they are at least as good as human drivers by some key metrics.
However, I've noticed that many work from the premise that self-driving cars are already safer than human drivers. That, or they'll work from the premise that self-driving cars will definitely be safer than human drivers within a reasonable timespan.
Worse, some assume that self-driving cars will dramatically, or nearly eliminate automobile accidents/deaths, where any brief skim over workplace casualties involving autonomous machinery would put that fantasy to rest.
For a group of tech workers that often overlap with self-described skeptics, it's an interesting blindspot to have.
You might have seen this internet classic:
https://www.youtube.com/watch?v=RjrEQaG5jPM [India Driving]
Driving in India is notoriously chaotic and if self-driving cars are ever deployed there in our lifetime, they will come with a safety guarantee about equal to a car driven by a blind dog with a missing paw. It would be impossible to safely navigate the streets there with anything less situationaly aware than a fully developed adult human brain.
The article above makes it sound like weather is the big problem for self-driving cars and once that's solved- wooohoo, we're on our way! It's far from that. Self-driving car AI is still incapable of reasoning about its environment, neither does it have any "undersanding" of it in any way, shape or form. Consequently it only works in very limited environments, in very limited conditions - of traffic, visibility, road quality etc.
(Paris) https://www.youtube.com/watch?v=lay8aZlsbB0
(Manhattan) https://www.youtube.com/watch?v=i8WiJgI3N4A
Not the OP, but like you say, the article is about a company who claims to have developed a technology that will solve self-driving cars' problems with rain.
People in industry make claims all the time. People in the sciences do, too. Just because someone makes a claim, doesn't mean it's true. It's only a claim.
You have a problem that is unconstrained, with infinite variables, where even a simple mistake can have catastrophic outcomes. Society itself may object to self driving cars for a ton of reasons, from safety to simply driving like a grandma and slowing everything and everyone down. The cost to develop this technology, plus the added cost of hardware to each car, will be enormous and is not obviously a cost savings over a $15 per hour human. If the self driving car is doing anything other than getting from A to B, you still need a human (or a human-like robot) to handle the unloading / delivery / whatever at the end.
Now, I’ve worked at companies with extremely talented and intelligent engineers, and something as constrained and seemingly simple as making a login form can take a long time to perfect - and no lives are at risk! Just imagine the challenges and requirements for building self driving cars. New hardware, software, real-time processing and analysis of tons of data, all to drive split-second decisions that can kill people if done incorrectly.
Huge challenge - huge risks - huge money - uncertain payoff. This is not something that will appear suddenly. If there aren’t convoys of self driving trucks operating in desert highways overnight, where it’s dry and straight and flat and no one else is there, then we aren’t going to see city taxis for a very long time.
It may not even be possible to solve without something radical like banning human drivers or inserting electronic nodes directly into our roads and infrastructure to aid autonomous vehicles. The existence of human drivers might make the problem simply impossible to solve in a way acceptable to society.
Software professionals are somewhat notorious for underestimating the difficulty of their projects. There's a massive amount of literature about that, proposing various techniques of mitigating this on a personal, team, or organization level.
That being said, I think in this particular case the problem was more of an overestimation of what the ability to work with highly dimensional data (as in ML algorithms) can give you (and underestimating the practical problems of deploying the ML-based systems). Basically, people tried running 30+ years old algorithms on GPUs while feeding them the ungodly amounts of data and realized that, with sufficient horse-power, they can make them (finally) work (as in: do something genuinely useful). This built up a lot of hype, of which self-driving cars are just one offshoot, I think.
Anyway, (if it's not evident from the above ;)) I find your arguments convincing and share your doubts. Self-driving cars are probably not impossible to achieve, but to get there we either need decades of research and progress or a couple of very high-profile breakthroughs in tech and theory. I won't hold my breath for neither :)
People are conditioned to accept that we will kill each other with cars every once and a while. I'm skeptical we will get to the point any time soon where the general public hears about a family killed by a self driving car and just shrugs it off. There will be intense pressure to get them off the road.
I'm not sure this is correct. I come from the optimal control world. The problem is not unconstrained and definitely does not have infinite variables (if you think in state-space, consider the state-space equation x' = f(x, u, theta). The x-space (state) is large but finite, and the u-space is fairly small -- steering, gear, brake, etc.). The x-space is also stochastic, and there are many observability issues.
That said, we've been designing control systems against the real world for a number of years now, and the key is not to model all the unknowns (because there will always be something you can never anticipate), but to model the known and safe path that the system can fall back to.
The problem is a complicated one, and it will take more than several attacks to make it work, but it is not by any means an impossible one if you break it down into the fundamentals.
Also, control systems have been used to control much more complex entities than just factories.
I agree it is an extraordinarily complex problem. But I also believe progress can be made to a point where it can be feasibly solved.
Frankly, it didn't change my opinion at all. As far as I can tell, Uber is a train wreck of a company. When I first read the story I was kind of surprised, until I found out it was an Uber car.
I’m curious how self driving focused groups are solving lane detection and managing surrounding traffic on snow covered roads. It’s not uncommon for a multi lane highway to be completely covered with snow and multiple tracks for different ‘lanes’ that overlap. It’s also not uncommon to see cars just making their own lanes. This occurs even after the snow has stopped flying.
Yep. As a Michigander, I totally agree. You see 6 to 12 lane roads reduced to "random" 2 lane roads when there is a good pileup of snow. And the snow drifts make it even worse!
With autonomous vehicles, you need to plan every situation in advance. Our current AI technologies can't improvise. It seems to me that current AI is basically just a decision tree with some neural networks sprinkled on top.
In general, I think, improvisation requires causal reasoning. It also often needs a broader knowledge than that needed for the nominal task (for example, when driving on a potholed road, one needs to have some understanding of how it affects the car's response, and sometimes of issues such as ground clearance.) Constraining the vehicle to avoid such issues would constrain the option to improvise.
Do you have any sources demonstrating examples of AI with true improvisational capabilities?
The last Waymo leak claims that it is more than just 'progress', but that 'rain is solved', incidentally: https://thelastdriverlicenseholder.com/2018/09/14/waymo-plan...
These sound like edge cases but there are places where they occur very frequently and local human drivers, for all their faults, do have a way of dealing with them.
I'd say that's pretty calculable, and has a solution, slow down to a speed that the tires can shed water.
Perhaps most telling about your qualification is the misplaced risk; the light precipitation with a little oil slick is likely more dangerous, mostly because you don't slow down for it.
Anecdotally, an even harder problem is when water exceeds 4", negotiating with traffic so you won't float away or get water in your intake.
In fact, they will have a lot more prior experience than any human driver in the world.
While it's quite fun if you pin up a cat dressed up in an elephant costume and surely amuses a few colleagues, a self-driving car is not something that should confuse an Elephant for a Cat.
ML is IMO not reliable enough for use in self-driving cars, in complicated situations the driver should take over.
Similarly, my 6 year old is also not reliable enough to pilot a car.
Both will change as they mature.
[0] I mean, there are going to be cases, but the number of people currently driving in horrific weather conditions because it would be inconvenient not to is too damn high.
On I95 in the DE-NY area, it's more like 60 mph in heavy rain. Or even in moderate snow. Everyone knows that they could never stop. But they're counting on others' inability to stop ;) But occasionally, there are massive chain-reaction pileups :(
Humans can and do drive safely in these conditions, however, and it's a valid concern to point out that Waymo's West Coast-centric models might not yet be trained to handle these cases.
(To be clear, I don't doubt that machines will eventually surpass humans in whatever weather conditions, if they haven't already.)
I've seen people do all sorts of foolish things. When the rain is so hard that I can't see more than 15 feet, I pull over and stop.
It's certainly not an unanswerable question. It's just quite hard to verify that they answer you have is a sufficiently correct one.
Why? It seems to me that estimating the confidence of a prediction by the quality of measurements is not an infrequent machine learning task. It'd be more complex for a driving algorithm than calculating the confidence interval of a linear regression, but it's in the same theme.
Self-driving cars would be putting lives at risk if they just pulled over and shutdown on every summer thunderstorm.
Can confirm, just spent 30 summer days in Florida. People don't pull over and stop, they just put on their "hazards" and slow down, sometimes to 25-30 MPH depending on the intensity of rain. This happens every single day.
If you're in the center lane and traffic is heavy on both sides of you, while your windshield (or camera) is overwhelmed with water, you need to keep moving with traffic for a bit -- or you get rear-ended.
https://en.wikipedia.org/wiki/List_of_countries_by_traffic-r...
Surprising that Waymo has figured out rain but not things like left turns. Different problem spaces I'd imagine, but still surprising.
The waymo cars have trouble with getting hit while making unprotected left turns because they drive very cautiously. Except for the time they hit a bus, because they somehow thought a bus would swerve to avoid them (although the then Google vehicle was being overly cautious and trying to avoid an inconsequential object in the road). All in all, I would rather the waymo style of overcaution and immobility than the Uber style of over confidence and ignoring of potential obstacles. However, I would currently still prefer mediocre to average human drivers on the road with me over a waymo.
From what I've read here and in the linked article, a Waymo car would be unable to pass a driving test here in Germany. Consequentially, they shouldn't/couldn't be allowed unsupervised on German roads, just like a learner driver. (In Germany driving schools are mandatory for a license and their cars are fitted with dual controls for the instructor).
> Also snow seems to be solvable, as there will probably similar patterns that the sensors get in their signals.
Real snow is not just a software problem. Fancy rear facing sensors near bumper? I hope they function through a couple inches of frozen muddy spray. Front windshield? Sometimes needs scraping. Lidar on roof? Does it work through the big heap of snow on it?
Still, for a wide range of weather conditions the autonomous car will just say a big NOPE and refuse to drive anywhere. Which is why it must always have manual controls. The autonomous car that could take me to the city through the worse conditions I experience every winter, won't have to take me in it because that car is probably clever enough to do my job as a software developer too.
But I guess when it's your product; you change your mind.
*This is not an attack to alkonaut.
Though I'm sure SV's real solution to this is to have perfect weather (ie. "can't reproduce; closing bug").
It seems that developers for self driving cars are dealing with a very different, and in some ways more difficult, environment than aircraft auto-pilots. While cars only have to deal with two dimensions instead of three and cars are still less complex machines compared to planes there are a lot more challenges. Unregulated and not controlled environment, traffic ranging from cars to trucks to motorbikes to bikes to pedestrians to playing children. Less room for sensors in the car, changing road condiitions, animals, you name it. Not a simple problem to solve.
And with a near perfect track record of similar systems in the aerospace sector the bar for acceptance hangs pretty high. Add to that the fact that pilots are trained in the use of these systems while the average joe / jane in a car is not. And the car should be able to drive without driver in the end.
So, I could imagine developers of self-driving cars could learn a thing or two from aerospace. But that is just my opinion.
Yes, sometimes you can pull over and sometimes you just have to put the blinkers on, stop in the traffic lane, and hope for the best--even though it can be very dangerous. One challenge is that, with full autonomy, the vehicle needs to be able to very quickly deal with a once-in-10-years scenario in the least bad way--both for its occupants and for any other traffic on the road.
This underlying misapprehension is at the root of the self-driving car hubris. Humans have remarkable perception of physical environments. The fact that electronic systems can beat humans handily on a few metrics has led to an entire industry built on the misconception that those few metrics are all that matter. But those were just the trivial bits.
The more sophisticated perceptual tasks--even the basic ones involved in the simplest self-driving task of traveling down a restricted access freeway--like correctly interpreting the behavior of other cars; interpreting whether the debris in the road constitutes a dangerous obstacle or something easily driven over; or even knowing which objects speeding toward you are moving and which are stationary--require more than just raw sensory input. There's a mountain of implicit cultural awareness embedded into all of those perceptual tasks that we haven't begun to form an idea how to represent in a software system. And again, this is the easy stuff!
I will basically guarantee you that no one working on self-driving cars thinks human perception is lackluster or an easily matched capability. Journos interpreting their work might impute that belief to self-driving engineers, but nobody could work on the problem for even the shortest amount of time without realizing that it is an incredibly hard problem.
In snow and ice, variables that are typically constant or static become dynamic. Formally dynamic variables now become multi-variable equations. For example: travelling down a hill normally requires dynamically factoring for the extra speed due to the incline. Under blizzard conditions you have to factor ice/snow volume, as well as the incline.
I don't want to share the road with self-driving cars in winter for a long time.
- Lane lines cease to be a thing, even if you can see them, you're better off not following them
- Stoplights also mostly become a guideline, it's sometimes safer to run them than spin out
- Signage or signalling may or may not be visible at all
- Other cars or pedestrians may or may not be visible at all even in daylight
- Your stopping distance can effectively change every few feet
- A foot of snow accumulation may be transferred from the roof of the car in front of you to the windshield of your car with little to no warning
- Your stopping distance can effectively change every few feet
- A foot of snow accumulation may be transferred from the roof of the car in front of you to the windshield of your car with little to no warning
The following are hard planning problems:
- Lane lines cease to be a thing, even if you can see them, you're better off not following them
- Stoplights also mostly become a guideline, it's sometimes safer to run them than spin out
But these are just damn hard and probably the true limiting factors:
- Signage or signalling may or may not be visible at all
- Other cars or pedestrians may or may not be visible at all even in daylight
But I'm not an expert on perception so... maybe that's my bias showing.
...northern Wisconsin gets its fair share of snow.
> ABS is usually the only thing they have
I think you've just agreed with me. ABS and TCS are exactly the systems I was referring to, and most of today's drivers have never encountered a car that doesn't at least have ABS.
I've driven a car without ABS or TCS in the snow. It's extremely difficult. Kind of similar to driving a stick shift for the first time. The idea is simple, but getting it into muscle memory enough that you're not constantly making mistakes is non-trivial.
Just like people could learn how to drive a stick shift.
But most people (at least in the US) don't know how to drive in a stick shift or how to navigate snow without ABS and TCS.
It's not that bad. You just dial back your risk taking to reflect the situation.
But in stopping, there is no such "spinning" or "digging". Instead, all the snow tires can do is roll or slide.
So, snow tires help a lot in going but not much in stopping.
Typically get into winter car wrecks from not stopping, not from not going.
So, once snow tires have you going, be darned careful about stopping -- the risk is not stopping, and snow tires aren't much help because the lugs don't get a chance to dig and throw snow.
That's not how snow tires work. Not at all. Independent of the design. And no, they don't 'stop' the way you describe either.
Finding #1: The main benefit of winter tires is improved tire adhesion, braking and cornering performance–not acceleration performance.
Finding #2: Winter tires provide improved traction on roads that are below 7 °C (45 °F) even when snow and ice are not present.
Finding #3: Stopping-distance performance of winter tires on packed snow is typically about 35% shorter than all-season tires and 50% shorter than summer tires.
Moreover, I was implicitly assuming that the stopping would be on snow, and not packed snow but, say, just snow as it fell although maybe a minute ago to several days ago.
For stopping on wet or slushy streets or just, say, 1-2 inches of snow, then, sure, "winter tires" not only help stopping but likely are better than snow tires.
But when my driveway has 8" of fresh snow, to get out I need snow tires, tire chains, or a tractor to pull me out. Then in stopping, say, from 30 MPH in such snow, nothing helps much except tire chains and there are questions about those.
Again, what is good about snow tires is that the big bumps on the tread let the tire, when it is spinning trying to drive the car, dig into the snow, kick it backwards out of the way, and let the rubber meet the road. For that, the aggressive tread, with the bumps, is essential. But the bumps work because the tire is spinning and digging, and that action won't work for stopping unless, say, going forward with, somehow, the drive wheels spinning in reverse.
The advantages of "winter tires" are how the tread cuts through water and thin layers of snow and has softer rubber with a better coefficient of friction when the rubber does meet the road.
Hell, I've been told I've not even to walk to work when it's snowing. ALthough saying that the last time it snowed here badly, I tried to walk to shop and fell over twice. Maybe my employer just knows me too well...
Driving in sbasic snow is a solved problem for humans in much of the world.
I live in Norway and beg to differ. We see temperatures down to -40 C in winter, and the kids love it. We don't have "snow days", it's literally not a thing here.
All you need is proper clothing, some seal fat for rubbing on your face, and cold weather experience among the adults (checking the kids faces for warning signs of frostbite, etc.).
95% of Norwegian babies and toddlers sleep outside in their prams (in sleeping bags) unless the temperature goes below -20 C. Then they come inside for sleeping, but will still be outside playing.
These are not things widely available to many Chicago children.
Can you elaborate? I don't understand this sentence at all.
Summarizing, pram is short for perambulator, which Americans call strollers. In Nordic countries, it's believed to be healthy for infants to sleep in fresh cold air. To the horror of outsiders, babies and toddlers in prams are commonly left parked outside of establishments in freezing weather while the parent is shopping or dining inside.
After 60cm snow on one night everyone is expected to be at school and work, on time, just like any other day. It should be noted that as a driver I also expect the roads to be bare at 7am even after that kind of night, and amazingly they almost always are. At least all major roads will be, but residential streets will not. https://static-cdn.sr.se/sida/images/109/b924f65c-8cc4-4332-...
Where I live, temperatures between 2C to -5C make the worst driving conditions (you're typically dealing with a mix of freezing rain, snow, ice, and slush).
Even though it's colder, -20C and just snowing isn't as bad to drive in.
Where there is little snow they don't have plows or salt/sand. 4cm of snow and you better stay home because the roads are covered in ice, and nobody can help you.
I have never experienced a "snow day", although there have been days with problematic traffic conditions due to snow. (Usually due to an experienced or unprepared driver going off the road on the first day of snow).
I live in an area that gets a lot of snow. The only 'safe' things for a human to do in snowy, icy, wet conditions are: slow down (a lot), leave a lot more following distance, stay home. I talk to a lot of people who seem to think that they are able to drive at full speed in the snow and complain about all of the 'out of town' drivers who slow down the roads because they 'don't know how to drive in the snow.' It's the people who make those kinds of complaints that frighten me the most - they are deluding themselves. I'd rather share the road with a computer.
Not version 1.
Factoring in all these variables is going to be a lot of work for a non-universal feature. I have no doubt they'll be included some day, but I doubt it'll be day 1.
EDIT: To be clear, I believe that dealing with ice and snow is an order of magnitude more difficult than driving under normal conditions. This is an 80-20 feature that many people won't need. I can't imagine holding the product back for additional development when just locking the feature based on a thermometer would suffice.
How do you meassure how slippery the road is? How much wind from the side? Will the wind blow me off the road because it has been polished by the wind just after those trees end, next to the lake? Will the car know to stop before even going because the car doesn't have winter tyres with good enough tracktion left?
A self-driving truck, does it know enough not to slow down in the uphill because it will start to glide sideways or backwards where the road leans too much to the side?
I get gray hair when I start thinking of all the problems that needs to be solved before we have self-driving cars.
One of the developer voes, am I wise enough to be humble about my limits of knowledge or do I know enough to be dangerous?
https://en.wikipedia.org/wiki/Traction_control_system
It's been a safety boon to have a computer limiting power to the wheels when they slip, humans are really bad at estimating how slippery the road is.
Estimating for the road ahead is a different problem, but there is data to feed into that estimate.
Actually, even with private ownership i'd not be surprised we'd end with region locked software, even if just for business reasons.
It's the same with a muddy and wet 4WD road.
The most raged filled responses I've even gotten to a comment on the Internet (20+ years) was when I called out aggressive snow drivers on Reddit. Despite being from, and living in the Northeast, the rage from fellow Northeasters whom I was accusing of driving like assholes was intense.
I recently drove several thousand miles this summer across many states (towing a trailer). It was only when I returned to the Northeast that I encountered aggressive, asshole drivers. It never occurred to me that maybe ... we're the baddies?
I think the previous poster was more about referencing recognition of patches of ice ahead and slowing down preemptively. This is a big issue in places with bridges, since ice can form on bridges but not the rest of the roads (since bridges aren't kept warm by the earth, the dirt under most roads acts like a heat battery). Also expecting when other drivers will slide. I've actually witnessed people getting, how would you put this, reverse-rear-ended by drivers trying to go up a hill and then sliding back down.
The challenge there of course is having the car learn to interpret the input of these sensors. But that doesn't seem like a unpassable hurdle. On the contrary, it would be pretty high on the list of things to tackle when aiming for driving in snowy weather.
(Also, as a data point, I think there was a guy who drove his Tesla with lights off, and it was able to stay within the lines)
Incidentally, if you've ever tried driving on snow/ice/mud in a 4x4 car with fully locking differentials, you have a completely mechanical solution that beats ESP.
And FWIW, Swedish insurance agency research pointed out last year that cars with permanent 4x4 and ESP were up to 40% more likely to get into severe accidents in bad conditions, because the driver does not get the "hey, this road is slippery, better slow down" experience, so they go too fast for the prevailing conditions.
http://feed.ne.cision.com/wpyfs/00/00/00/00/00/3E/5A/A2/wkr0...
I learned to drive on ice covered dirt roads, many times I would have liked to have the ability to asymmetrically brake. A simple 2nd order gimble would be awesome. Having a single symmetric pedal dumbs it down way too much to make a pre-programmed car vs human comparison.
Same thing goes for torque distribution.
It's generated dogma that pre-programmed devices can perform better than biological systems, so I bet my suggestion will be met with incredulity bordering on "y humans cant handle x".
Yes, I have been in that situation too.
As for brakes.. ABS is very bad at handling hard snow with a drizzle of fresh snow on top. It's nearly impossible to stop. I've had to actually turn off the engine in order to stop the car when driving (slowly) down a hill from a hotel in a certain snowy area, in order not to end up on the main road in front. That was with various rental cars (yes, I go to that place regularly in my job). The first time this happened I thought here was a technical problem with the car. So I asked around. Same experience from others. The cars vary, but they're all much worse than fully manual brakes, in certain conditions.
And then there was the time there was a reindeer (yes, a reindeer) running along the road, zig-zagging in front of the car. Snow conditions as described above. ABS refused to brake. Didn't want to switch off the engine (busy steering), so after about fifteen seconds of this I drove the car off the road instead, safely into a lot of snow.
And people want computers to drive cars, when computers can't even be used safely for brakes?
In any case, I've yet to hear about a self-driving car which can predict that another driver will get into problems in a few seconds unless I slow down or increase the speed, in order to allow a third car to adjust. This is something I see or do nearly daily, on my commute.
Its a trade off.
If I mudulate my brake lightly while threshold braking, I can usually stop 15% faster than similar application of ABS. However, when I mindlessly slam on the brakes, ABS wins by a margin of 20%.
This is not the full picture however, and better driving skills alone do not negate some abs advantage. The main one being control in Turns while braking on poor surface. Without a knack for sliding sideways around corners, ABS was clearly superior when braking during a change of heading.
That all said, I run without ABS, as it suits my personal driving abilities. This would be a mistake for most people. Fortunate for me, most performance cars afford me the option, while some others consider it a critical system, and will not even activate the power steering unless detected as working. As if steering somehow became less critical without abs.
Modern car engineering is still trying to turn a tool into an appliance, and it feels like a severe mismatch of values to me.
HAL: None whatsoever, Frank. Quite honestly, I wouldn't worry myself about that.
There's a really short description for self-driving cars -- hyperhype.
And I'm in two wheel drive and not 4 wheel drive because the extra torque it takes to turn the machinery for 4 wheel drive is too much for the poor traction.
I stop at the top of the driveway, make sure I'm in 2 wheel drive, hold the transmission selector so that I can get into neutral quickly and easily, just slightly get the car rolling, go for neutral, mostly just let the car roll down at most just touching the brakes very lightly or not at all, let the car get to the level, bottom of the driveway and a start on the back yard, and then stop the car, go into reverse, backup a little, and then take a left turn into the garage.
Losing directional control and having the car turn sideways and roll over is a big risk. The driveway is not very long, but it's steep.
Here's an interesting bit of research on filtering out snow noise from a LIDAR point cloud:
http://wavelab.uwaterloo.ca/?weblizar_portfolio=real-time-fi...
(You can have a map with lane markings and localize on that map, though)
It seems to me that a lot of people commenting here are a bit obsessed about the edge cases, and if it has edge cases then it’s not a “true” self driving car. I disagree. I think a car that can self drive in some situations would still be amazing.
Does that mean they were useless? Absolutely not!
Lately the mindset seems to be unless a product is all things to all people at all times then it's useless.
In reality a product just has to fit a category really well and it will sell like hotcakes. Then it will evolve in iterations.
Much of the dream of self-driving is for no human driver there, for taxi cabs, school buses, 18 wheel trucks, local deliveries from pizza or Chinese carryout, USPS, UPS, or FedEx, etc. For that, for current technology, for current traffic, on current roads, there's no hope at all because the edge cases are way too common in practice (can't put up with mean time to destroying an 18 wheel truck of six months -- 5 million miles is more like it) and require full, wide awake, sober, mature human intelligence with full ability at reading, talking, understanding, natural language understanding, hand signals, flag signals, tough to read road signs, etc. Edge cases.
A self driving car that can't drive in all conditions should be positioned instead as a driver assistant that can't assist in all conditions. If the expectation is that the human driver is doing all the work, but sometimes the computer will help keep the lane, or with emergency braking, it's OK for that to not happen in rain or snow -- it was always the human driver's responsibility, the computer will help when it can.
Also, that time-limited "hold the wheel" mode would typically be engaged exactly for those times when the human driver is not at their best (tempted by communication devices, tempted by the food basket or just wrestling with navigation), so achieving better than human would be relatively easy.
So, the intelligence of a mouse, crow, parrot, dog, cat, monkey, elephant, dolphin, orca, etc. just is not enough.
So, self driving cars requires essentially full AI for driving and nearly everything else in life. So, the self driving car problem is no easier than the full AI problem.
Sure, there can be some special cases that are easier -- trucks in a huge, open pit copper mine, a tractor on a huge, flat farm in Missouri, some military battlefield situations, and the public roads if heavily reengineer them with lots of essentially electronic tracks, etc., uh, right, on roads in perfect condition, no big objects falling off trucks, no drunks, no tire blowouts, dry weather, daylight, perfect visibility, no alarms from an extensive monitoring system, and no rain, sleet, or snow. Then, sure, a "self-driving car"!!!!
Still in the end it requires a combination of changes to how we mark roadways, indicate construction, detours and the like. Think of it like the ADA but for cars.
Safer is better, no matter how we get there.
Not really true. It'd be safer to just get rid of high-speed vehicles in the first place, but that would be disastrous.
https://thelastdriverlicenseholder.com/2018/09/14/waymo-plan...
seeing the fun my dog is having chasing the birds, i'd suggest that a Boston Dynamics dog would jump out of the car, scare the birds away, jump back (and continue to drive while watching electric sheep funny videos :)
Can any self driving car read parking signs because it's not like there is some API for when/where you can park or drop passengers off.
As someone who had to relearn the rules of the road for a new country - there is a lot that's unsaid.
Together with some regulations, of course. Though, this already has to be regulated. Not just anyone can put up a "only park during certain times" sign. (Right?)
But worse: RFID makes for duplicate signage. What if the visible and data signage differ?
And in this case, any discrepancy would be reported back by any car passing by, so the error could be corrected quickly.
https://www.youtube.com/watch?v=t7jxensSdhE
https://www.youtube.com/watch?v=IMlz5tAKmUs
They've been working on this problem for a very long time. Which explains the progress they've made. I don't know how people can say this is an unsolvable problem when some high end research applications have made such progress.
Plus there's Microsoft's 2016 announcement with Ford about algorithms that can work within snow/rain by identifying raindrops and snowflakes then 'ignoring' them:
https://qz.com/637509/driverless-cars-have-a-new-way-to-navi...
Edit: Grammar & some terminology
[ ] Let self driving cars kill people during development but try to make it up by saving thousands of lives after they are perfected. (Uber)
[ ] Make self driving cars extremely safe from the start but more people end up dying from manual driving because development takes longer. (Waymo)
Most safety problems in driving can be solved by slowing down.
The problem with AI is all those weird, little edge cases that humans can reason through -- for example: if there's a deer next to the road, then I'll slow down, even if it's not on the road yet. I've known many people who have hit a deer when it spontaneously jumps into traffic. Or something like: someone's not quite staying in their own lane, so I have to be careful when I pass them.
I doubt that this is a use-case that legal, for-profit companies will be pursuing.
For the other use-cases, you can just say: Manual driving only.
Here's a restatement without all those corpses: In the event that an accident appears unavoidable, the computer should do everything within it's power to stop the car immediately. (This is already a given: safety first.)
"If there are multiple ways of doing the above, what variables should the computer optimize for - stopping distance alone, or stopping so that it doesn't immediately get rammed from behind?" There's your trolley problem again, just restated so it doesn't appear so offensive: in both cases, the occupants of the vehicle are in danger, as are the occupants of nearby vehicles. Now is the interest in the rabbit-hole clearer? The problem doesn't go away just because it's inconvenient to solve...
We tell humans not to swerve for squirrels, because it's a great way to end up dead, so the computer should get the same instruction. Slow down as you can, the prime directive is to maintain control.
In other words, there is always a balance between safety and usefulness, and the question of "is this the right choice" is always upon us, whether we want it or not. You can never have certainty anyway, what the software is doing is maximizing on some reward function. The TP also asks "is there even a moral component to this?" It seems there is, from the range of emotions this conjures up.
(btw "do not swerve to avoid unknown objects" has directly caused at least 1 dead person - Elaine Herzberg - so that's a really unfortunate maxim for illustrating your point: the car has also an obligation not to be a danger to others. "We just maintain control and everything else be damned" is easy to explain in court, true: IANAL, but sometimes you need more than a simple explanation to avoid being convicted.)
Elaine is not a great example because not only did the car choose not to swerve, it also chose not to even try to brake. And on top of that, a human would have seen her from a lot farther away.
And the larger point stands: "protect occupants, ignore outsiders" is a choice in the TP, always choosing the same strategy is not "TP is irrelevant."
Lidar's record much less information than than 2 optical lenses (eyes). You got distance in 1, 2, or 2+1 dimensions, and intensity. Eyes can detect distance, wavelength, reflections, refraction, contrast, brightness etc. All with the assistance of the brain of course. If today's computers and software had the pattern matching capabilities of even a human child's brain (reasoning aside), a few moderate resolution video cameras would be all that is necessary for self driving.
But we're not there yet so we employ sensors with limited detection that is simpler to interpret programmatically.
I think even humans would have trouble interpreting the mess of lidar distance graphs in a rain/snow storm, though they'd still do better than current computers.
Perhaps it is that the rain drop refracts the beam to all sorts of other objects which can make it difficult to tell which actually has the longest Z-distance return?
In comparison, basically no self driving car could safely drive 100 ft during a snowstorm.
Getting self driving cars to be as good as humans during harsh weather (which is not that good I'll give you that) will take 10 years in my opinion (I say this as a grad student in mobile robotics).
Autonomous cars are already safer on ice. being able to individually apply abs to each wheel in micro second speeds of detection of change in terrain is far more accurate than any human.
As far as snowfall, for the vast majority of places, hi def maps and semantic knowledge of signage already exists, so as long as the car can localize, it can obey snow covered signs, and google and others are already working on filtering out weather "noise".
The problem isn't as intractable as it seems
Rough localization in a snowstorm (as in good enough to use prior knowledge of signs) or rain shouldn't be too hard. GPS still works in bad weather and lidar localization (scan-to-map matching) should be hampered but still usuable.
Seasoned New England drivers are idiots. Source: Am seasoned New England driver.
The idea that we can quantify all of the subtle little ways the human brain instinctively has gotten so good at driving cars is supreme arrogance.
Things like burying a wire that can be followed, embedding RFID devices in the pavement (along with an online database that gives their precise location and any updates on road conditions), etc.
It would be conceptually similar to all the ground based systems that guide aircraft.
Of course, these would be deployed only on high value roads like freeways first, and you could at least let the autodrive work on the freeways, and do the last mile yourself.
Where I learned to drive, most humans could 'see' black ice by predicting where water will pool, the recent weather, and subtle clues that abound, but are not easily expressable. Any out-of-town'er would surely be spun out or driving 10mph, white-knuckled, aghast at the "maniacs" flying by at normal speeds. While in my college city, I had to relearn everything, as it was rainy, and the drivers hyper agressive, and 15mph faster. The same model would struggle to encompass both modalities.
Until AEB stops accidentally triggering constantly on colleagues cars, I have zero interest in trusting my life to an algo connected to a 300hp steel cage.
It certainly feels like humans become improved drivers through various weather / road condition experiences, but I sometimes find myself wondering whether my gut instincts would align with a scientific approach.
Its a skill. You get better with practice. The person going slow probably just hasn’t had enough practice yet with the conditions. Going slow is probably a good choice for them. Am I a better driver overall for knowing how to handle ice? Probably not, but I might be the best person to do the driving under those conditions if everyone else in the car is from warmer climes.
I remember standing in line at Minneapolis rental car counter one night when it was snowing an inch an hour with 4 inches on the ground already. Agent to the guys in front of me: “There are chains in the trunk if you need them.” Customer with heavy southern accent: “Chains? What the heck are chains?” I wonder to this day if those guys made it to their hotel.
Same deal for where the rain will be bad, where I might need to worry about hydroplaning, etc.
Thats literally all I change, and I have done fine all my life
It was described in the drivers ed manuals when I got my license, but it is a skill that has to become muscle memory habit (i.e., you don't have time to think "which way do I turn the wheel and how much", you have to "just do"). Without any practice (which was way easier in the 80's with rear wheel drive vehicles) you never got the "touch" down to just turn the wheel the correct direction and amount to recover when the car does start to skid.
One cannot discont the psychological factor of feeling 'in control' when the control decides if you or others might die.
They also tend to appreciate the importance of winter tires more.
This is anecdotal, but based on a rather large sample of people I know who moved to a nicer climate.
Personally, i'd say there's a pecking order from awesome to awful for the median driver in snow as follows: Buffalo/Vermont natives, NYC/Boston, DC Metro/South, Florida/Texas.
Likewise, if you're from the Northeast, driving in rain in the southwest is surprisingly hazardous.
We, collectively, not just the US, have a problem with too many cars and instead of reducing said number of cars we want to make them self crashing?
Instead of investing heavily on public transit systems we keep feeding this fable that somehow, in the near future, cars driven by computers will do at least as good a job of driving as sentient, sober people.
I can't get my smartphone to understand my language and do simple things by voice commands and yet we seem to think that there's (nearly) available technology to make a car drive by itself...
These companies must not exist in the real world, that's the only possible explanation.
Google and Uber believe they can make an insane amount of money with the transition to automated driving. This is being sold to the public as the solution to a public safety crisis with the narrative that human controlled cars are death machines which need to be taken off the roads as soon as possible.
I don't know what ticks me off the most, people believing that these companies can easily solve the automated driving problem or these companies stubbornly pushing for it.
I'm just utterly surprised by how so many people think this is a fixable problem in the near future, it's not.
How long until people realize that?
The next AI winter won't come soon enough.
Once there Uber self driving cars will then be both cheaper, safer and more reliable than your own car. And you don't need to park them. Only those with special requirements (ie professionals that need to store equipement in their car) need to have their own, the rest of us can use a taxi service.
Furthermore, when human driven cars are removed from the roads, road capacity could increase by 50-100% or more, as the distance between cars (in both dimensions) can be safely reduced.
Also, you will not need to pick up your children/parents/others that cannot drive. For instance, if you drive your kids to school, you likely have to drive both ways, doubling traffic (and wasting your time).
For commuters, self driving cars also lend themselves were well for ride sharing, having cars that function like full office spaces or entertainment sources, meditation rooms or whatever you can think of. If all cars are self driving, the speeds will could be regular, and uncomfortable acceleration/deceleration avoided.
Finally, the tech going into this is likely to be generalizable to other kinds of machinery, such as mining machinery, construction machinery, automated chefs, as well as the functions in manufacturing that have not already been automated.
I think the potential of the end state should be quite clear. The total economic potential is huge. Anyone able to attain near monopoly for one of the technologies going into this (software, sensors, traffic control centers, etc) stand to be the next Google.
That said, I think it is unlikely that it will take off in full until between 2025 and 2035. Now, we are in a situation similar to the .com era around 1998. If history repeats itself, 2023 may be a great year to invest in robotics startups, provided we see a bubble bursting.
If only I could see the same enthusiasm for say, improving public transit range, infrastructure and operational costs... nope, auto-cars will save us.
Take a look at Germany, they've just launched the world's first hydrogen-powered train.
As a side note:
The total economic potential is huge. Anyone able to attain near monopoly for one of the technologies [...]
Economic potential and monopolies don't mix very well, unless you're one of the shareholders of said monopoly.
On monopolies: Look at the big companies coming out of the internet revolution. Google, Facebook, Amazon, eBay, Uber and the rest are all near-monopolies in their core segments. As are older giants, such as Microsoft and Apple.
I'm not saying it is good, I'm just saying that the lesson from the previous wave, is that the winner takes it all. And as investors have learned this, they don't want to be late for the next party, hence the hype.
Of course, the risk that the bubble will burst at least once before the actual party, is pretty high.