We already know that humans don’t handle many cases very well. Deep water. Driving too fast on ice. Driving too close in heavy fog. And there is clearly a long tail of bizarre traffic situations which won’t be handled until they occur.
We don’t get our licenses fully prepared to handle all situations. We learn through experience. Until your system no 17-year-old would be allowed to drive. The difference is that we’ll have just one 17 year old learning, not a country full of them. And ours won’t drink.
The Wiki entry on limitations says Google plans to fix all the things you mentioned by 2020. Most of them are not inherently difficult. If the issue was whether the cars will be ready next year, this would be important. Not that long ago cars couldn’t make it through the desert by themselves. Now they drive on busy freeways. And yes, the Google people are smart enough to know about snow and even, despite the drought, rain.
Don’t have resource to map? Ever see street view? Yeah, it is going to be a while before county roads in the deep woods get mapped. Roads with 95% of traffic, not much. And I don’t think the car with no steering wheel will be released, at least not for a long time. It makes a point but it would kill sales, and the auto companies who will really sell this car wouldn’t go for it.
I wonder what the accident rate is for the roads the cars travel. Based on the traffic reports I see a lot more than 11 in a million miles, I suspect. And remember, the Google car is still a teenager. Or maybe it is a distraction to drivers near it. The computer itself seems to have caused zero accidents, perhaps a better measure.
Soon after my daughter started to drive she was stopped on an off ramp when a car coming off the freeway hit another car who hit her. (She did not hit the car in front of her.) I didn’t ban her from driving. Maybe you shouldn’t blame the computer for another car hitting it at a light.
Yes, those saying cars will be commercially available in a year or two don’t know what the problems are. They are just as wrong as those saying they won’t be available in 20 years.
OK if we go with production cars, there is a system today, in use and proving itself, that will take priority over the driver and override him/her. The trend is here already, the computer is better at certain situations and will continue to get better, and also as computers are going this will happen very very rapidly.
The person who instructed both the previous and the current head of the Google program doesn’t think they’re likely to conquer the AI portions in his lifetime, and he doesn’t have anyone to satisfy but himself. Let’s just say the level of difficulty in that is not defined at the moment, ok? Maybe they’ll solve it, maybe they won’t. I’d wager that right now, they don’t have a firm plan on how they’re going to solve most of them.
Look at the article quoted before. To make the map used by the car requires human intervention. It’s a problem far beyond street view. The same article indicates that pretty much every auto maker besides Google (who still haven’t made a production auto, ever) thinks that the car without the option of driver control is sci-fi. I hate to repeat myself, but it appears that you are not aware of the problems inherent in developing a self-driving car, or you are ignoring them. Maybe you are on the side of the near future, but you haven’t explained how that might happen. There’s not even a theoretical framework that is workable.
Seriously? This has already been addressed The average driver currently has around 4.19 accidents per million miles driven, irrespective of who was at fault(this was already linked to before, see page 4)
I wouldn’t, but I also wouldn’t claim it as a triumph of the software if that was the only accident it had gotten in while it was a teenager. I had no accidents that I could be assigned fault for while I was a teenager, or even before I was 30. I could beat the Google car’s record during our first 750K miles (only 8 accidents, none my fault). I participated in semi-organized street racing during that time. No one with a modicum of a brain would would have called me safe. Considering that the Google car’s record has been racked up under such limited circumstances, I can only regard it as an anecdote in the development of the driverless car. I don’t even think of it as truly self-driving yet.
Meh, I’ll probably fall into a foolish debate about “Which is more improbable:FTL or Strong AI?”, but we don’t know which of those is more likely to be achievable at the moment. However, only one of them is necessary for a self-driving car that can handle the variety of situations an average human can.
Yes, and I am saying the idea of automatic braking (which is what all of those are) is good, provided it can be disabled or overridden when something smarter or more aware than it decides it is prudent. The most desirable function is that it would operate seamlessly, and allow the human to take control via the standard controls when necessary, such as my WRX appears to behave in regard to traction control.
Have you ever seen what towns do when they add stop signs to existing roads? They put up multiple other stop signs and “Warning STOP Sign Ahead” signs. If they didn’t, people would blow through the new signs 100 times a day.
I’ll also point out that your 4.2 accidents per million miles is “police reported” accidents. The next step is to estimate how many accidents go unreported. It would not surprise me in the least if a sizable number of minor accidents go unreported. Accidents that Google would absolutely report as part of their testing process.
Over 40 years ago I took an AI class at MIT, where AI was going to be coming any day now. I’m not a fan of strong AI - we have made almost no progress on it since 1960.
However the specific areas of AI which we studied then have pretty much all happened. They include voice recognition, solving mathematical equations, plotting routes between two places, and chess. I know you think a self-driving car needs strong AI, but I dispute that. On my commute navigating on 880 requires a lot more attention and smarts than stopping at stop lights on the way to it.
(Solution to the stoplight problem: new lights send out a radio signal saying if they are red or green or yellow and which direction. Cars read them. Problem solved cheaply. Old lights get mapped.)
A self-driving car exists in a limited domain. You care about cars along the curb and about any cars coming out of driveways, you don’t care if someone is sunbathing on the lawn.
I’ve been developing complex algorithms for real applications for a long time, some of which were considered practically impossible. At this point I can pretty much sense what is feasible. Strong AI isn’t. Driving better than people is.
As for the accident rate, who is the better driver - the guy who causes his 5 accidents or the guy who got hit by others five times? The Google car is the former. I was not asking about averages everywhere - just in the Bay Area, by the way. With roads that scare our traffic columnist. Now, if the car under autonomous control caused no accidents, if most cars were under autonomous control the accident rate would fall, wouldn’t it?
The reporting rate is a good point. Some guy hit me, lightly, at a traffic light. No damage to either car. It was an accident, but we both went on our way and that accident did not become part of the statistics.
I definitely agree that any car sold is not going to be self-driving only. Being able to take over extends the places you can drive it, at least in the beginning, and will make people feel more comfortable with the car.
If Google sold a car for a reasonable price that could do the things their cars already do I’d buy one in a flash, since it would relieve me of driving for 95% of my commute. And in 20 years when I am 83 I’ll buy one (if not before) since the car will be safer than I would be. The Senior Wheels people are really going to suffer.
Where I live we get standing snow on the ground for a few days every 8-10 years. In those few days, car accident rates skyrocket. even humans can suck at driving in snow. small snowdrifts do present a unique challenge that I agree must be addressed for wide scale deployment. That said many of the challenges in snow driving can be addressed with an “ice and snow mode” that would reduce acceleration, braking rates, and cornering speeds to account for reduced traction. My first thought is if the car can manage a large speedbump it should handle smaller snowdrifts. I would imagine some combination of radar/visual/ir could fairly accurately determine that the large white obstacle is a snowdrift as radar is already used for measuring ice and snow depth. Probably not an unassailable problem especially if my forementioned “snow mode” is active therefore anticipating minor snow drifting. If drift density can be measured by radar it could also probably detect the difference between a drift and another obstacle covered by snow.
Not to mention - what if the cars have autonomous mode and person controlled mode?
For those situations - where there is no snow, then the car drives. When the roads detoriate to the point where the car can’t drive - it stops in the nearest parking lot and the human takes over. (or it doesn’t even start)
Right NOW BMW already has live traffic reporting, sent directly to the car, if most / a lot of cars are autonomous - how hard is it to have accurate information for the vast majority of built up areas?
I also don’t see a problem with mapping - once cars can share mapping information to a central server - unmapped areas will be mapped pretty damn fast.
Currently I live in Singapore.
My estimate would be that easily 50% of cars never leave this city.
A autonomous car would absolutely take off here - we don’t get snow, or ice, there are no “off road” situations to deal with, the roads could be mapped to within millimetres wihin the first month of operation for autonomous cars.
Our situation here may well be unique -
But I can absolutely see where there will be a great many people in other places where their personal cars are used exactly like the situation we have here in Singapore - only driven in well known city areas, no snow, etc etc.
There WILL be a market for it, the cars WILL be able to handle it and it IS coming.
Most of the arguments against self driving cars are of the sort
But it doesn’t handle a traffic diversion through a field
It can handle a 12" snow drift
My contention is basically “SO WHAT?” - there is enough of a market in the areas where an autonomous car will excel, and be safer to make it a reality sooner rather than later - even if for some places and some people an autonomous car will fail spectacularly.
I dont see 100% autonomous driving for at least 10 years after they hit the street. many if not most of us dont see serious driving challenges on a daily basis. Im not going to see it as a meaningful failure if I have to give the car a nudge once in a while or dealing with traffic cones/construction or the occasional weirdness. A car that can handle 80% of all driving situations is going to get the vast majority of drivers through their daily driving with little if any human intervention. if Taking over control for 2 min to navigate a cone lane redirect is seen as a major failure we are probably hearing from people who are rarely happy with anything.
That’s pretty much precisely my position. Additionally, I’m saying that self-driving isn’t going to be mandatory any time soon, but we’ll see an adaptation of progressively better driver’s aids.
I’ll stand by my claim that a level 4 (autonomous) car would require strong AI to be very useful. It’s going to have to at least be able to differentiate between the wide variety of objects that can end up in the road, and which ones it can run over. If it can’t, it’s going to be a petty big hazard/obstruction itself, and I don’t see how you’re going to do that without strong AI.
Please keep up. The car cannot handle SNOW. Not a drift, not a blizzard, snow. Snow falling from the sky and covering the ground with less than 1" is out of its design parameters. When its falling, it blinds the sensors it depends on to find anything. When you do get drifts, it most likely can’t orient itself on the map, because it won’t be able to line that map up with its sensor readings. They do not have plans to attempt to conquer snow any time soon.
And the map is a big deal to maintain. Google deciding to do it that way lets them do some impressive stuff, but it’s a giant weakness for the car, IMHO. Even the specialized car Google sends out to make the maps for the car (which is an entirely different beast than the Google Maps car), can’t make the map without a human going through the data and massaging it. If Google’s version of the car is going to become practical outside of their sandbox, they will have to figure out a way to automate the map, or make it not necessary.
ETA: And again, the car isn’t even tested in heavy rains yet, due to safety concerns. Get much rain in Singapore, do you?
I’m fine with automated cars crusing down the highway. I argue that it MUST allow the human to take complete control whenever a human wants to. Some folks seem to think that a computer is ALWAYS going to handle anything better than a human.
And here is an obvious, flag waving, red light disconnect. A 12" snow drift? Bwahaahahahh.
That’s contradictory. If you can’t trust a car not to run over a person in the road, you can’t trust it to drive itself at all. That’s far different from saying that the car can’t find its way in the deep woods or in a blizzard, and should be manually driven under those conditions.
And the basic recognition of 99.9% of the things found in the road doesn’t require anything like strong AI. Most people make snap judgements about whether the obstruction looks hard or soft. It’s a plastic bag - you don’t care how big it is (within limits) or what store it came from.
Not that people do this perfectly. Mr. Roadshow just ran a column about a poor couple who totaled their beloved car in a pothole. (We grow them big out here.) I bet even your basic Google car would do better than that.
So, how’s the car going to ID the object in a ‘snap judgment’. The human uses it’s brain’s ability to recognize objects using their lifetime of experience. The computer doesn’t have that benefit, and I don’t see how it’s going to get it without strong AI. You haven’t presented an alternative method. If you know of one, you have been wasting your time writing debugging code.
As far as that being contradictory: hey, I’m just describing the car as it is. It can do limited self-driving, but cannot differentiate between common objects. If that’s contradictory, call reality, not me.
The Google car can’t avoid potholes, either. It can slow down for them, but it’s going to hit it. If the pothole can total a car, I don’t know if the Google car would do better.
You identify things through a lifetime of experience. Yet you can take a picture of something on your smartphone and have Google identify it. And the domain of things laying on the road is much smaller.
I’m not talking about what is available now - though the cars that are out there handle things pretty well already. I’m talking about the cars when they are ready for sale. Like in any reasonable research type project, you get the core stuff working first and then you develop additional features, like the ability to drive in snow.
Actually, your link says it sees potholes, but slows down instead of avoiding them. There are so many potholes around here that this is the right thing to do. Swerving can cause accidents. There is an immense backlog of road maintenance here, so the average driver spends in the multiple hundreds of dollars a year on car repairs due to bad streets. Until the Google car can fly, it will run over pot holes too.
My dog says that offing squirrels should go in the plus column, by the way.
ya wanna talk potholes? Heh. I can move left to right on my gravel road and miss the majority of them. There is no traffic so that’s not an issue.
And what about snow? It will have to measure water content to be able to tell if it is 2 feet of powder, which is no problem, or 15" of heavy snow that can be an issue.
I take issue with those that think that a COMPLETLEY autonomous car is a good idea. And that the driver can never take control.
One poster said that it would be fine if the car said it was unsafe to drive and refuse to move. Another said that there should be no way that the ‘passenger’ should be able to take control of the car. Another thinks that 12" of snow is a snow drift.
Sure, there are only 10s of millions that deal with this (and many other situations - wanna go to the hospital?)
It gives me a little insight to how sanitized some folks lives are.
This all depends on how you want to use the car. For example most families are multi-car households. With an autonomous car, it could first take the husband to work, come back home and take the wife to work–and likewise in the evening. So the family only needs one car. Or you could have a fleet of autonomous Uber cars.
