The conclusion about map generation is a bit naive. Yes, if google's self-driving car requires them, and google can't currently produce them at scale, then maybe google's current solution isn't going to explode onto the marketplace. But could the current solution be used in certain cities? (start with San Francisco, then expand city by city as the demand exists) Sure. Could google come up with a different solution? Sure.
There are other companies working on self-driving cars - BMW, Audi, Tesla, others. The darpa grand challenges in 2005 and 2007 didn't rely on this kind of detailed map data.
Will self-driving cars ever be able to handle all possible driving situations? Probably not. But humans can't handle all driving situations either. A more relevant question is whether self-driving cars will be better than human drivers. It's silly to say that we need strong AI for that, since we've already seen several prototype systems that do better than humans in many situations without strong AI.
The other thing that this conclusion ignores is what the changes to the road system might be. If everyone gets semi-autonomous cars for commuting, and they work fine except for one intersection where people need to pay attention and negotiate it themselves, then there will be pressure to change the intersection. Maybe besides the carpool lane, you might eventually get autonomous lanes. It's a bit chicken-and-egg, but if there is a system that is useful enough in some situations for people to buy, then it will progress from there.
Yep. To take a step back, the whole point of tarmac roads is that they are designed to be easy to drive on compared to cross country, footpaths, goat-tracks, wagon-trails etc. It's not hard to see a tipping point where roads are designed for robot cars not human-driven ones.
Also economies of scale take over. If 100 robot cars an hour drive down a road, the details of it can and should be mapped in real time and shared among the robots.
Google has almost 1 million miles on the road with its self-driving cars. I would be shocked if they don't add or improve their mapping data with each mile those vehicles travel.
I think the author is also missing the implications of static mapping being only a partial paradigm in the car as a whole.
While I can't speak for the specific project, I imagine in addition to static/procedural mapping, there are generalized machine-learning algorithms & gradients which take into account arbitrary weights on context (traffic laws, potential loss of life, traction & other situational data, avoidance costs, actions of surrounding drivers, etc) and optimally, produce a decision that is as close to a human reaction as possible, while boosting efficiency and minimizing for injury.
And I imagine the resources are available to test the edge cases in simulated conditions more extreme (though perhaps less variable) than those encountered on public roads.
It may be another 5 years, but the solution is simply one of technological advancement--better static routines combined with more efficient behavioral models that produce better outcomes than the reaction of a human nervous system.
(And this is somewhat disheartening for an admitted 'petrolhead,' but innovation is an unstoppable force)
The German automakers have been working on a sliding scale of automation, with driver-assist systems sort of transitioning into full automation sometime in the near future. (There's a nice table on page 166 of the 2014 annual report of the VDA, the German automobile manufacturer's association: http://www.vda.de/en/downloads/1225/?PHPSESSID=fvu1lhmadgb23...) (Note the incredibly good translation job - not just anybody could have managed that.)
The main point is still valid. In an emergency situation, as long as a self driving car has to choose between hitting a ball and a kid chasing after it, or a woman and a shopping cart, or a laid down motorcycle and its rider, and not know which is which, there is no way these cars can be deployed in mass. And this is an extremely hard problem.
I for one don't trust myself with that judgement call either (given the expected very short timeframe you have available).
For example, avoiding the kid that appears out of no where might crash me into an upcoming car, killing that driver while the kid in the corner of my eye turned out to be a dog or in fact stopped/retreated just in time to prevent a collision in the first place. Your fastest reflexes bypass a big part of your brain.
As long as the cars drive safer, on average, than we do (and I think that's achievable), rationally we should hand over the controls (or we should find a way to combine both and end up even safer).
One might make the same argument about the cars we have now. If we cannot guarantee that there will be no deadly accidents, they cannot be deployed in such great numbers. Yet, car accidents are a major cause of death, and almost everyone drives one regularly.
In an emergency situation, just stop. Don't hit the ball or the kid. Let the human take over, or wait for the emergency to pass. The car will break faster than a human could. If the car can't drive itself safely in a given situation, make humans drive.
There are valid points to be made about liability and legal concerns, increases in road congestion, etc. But technical considerations are not really an issue.
You don't need to completely solve all potential problems before entering the market. You just need to provide enough value.
If you've got an hour commute over highways every day, even being able to do highway driving automatically will be a huge win.
That's usually the right answer, but you have to be careful. For instance, you say the car will break faster than a human could. But if it does, the human-driven car right behind it is in trouble.
Yes, you'd want to have some sort of sensors behind the car as well, so that if it's possible to avoid hitting the object and avoid being hit by the car that you would brake at that speed. If it's not possible, well, it's still an improvement over a human driver.
People "often" (for made up values of often) hit kids instead of balls bouncing into the street, as their eyes are drawn to the ball and never notice the kids scampering after it. A machine can be given enough resources to have the scan of a tree full of owls - no single driver can.
Coming from Europe, I can see self driving cars work in the USA, your roads are just so big and... well "easy" to navigate. Now come to Europe. Narrow winding roads, no grid system in cities, lots one way traffic, uneven 'brick' roads, blocked often by wrongly parked cars, thousands of bicycles.
I have a hard time navigating here myself. A lot has to change before the automatons take of the streets :(
Well road testing starts January 2015 in the UK, and in Gothenburg from 2017 [0], presumably many other European countries have similar plans so we'll see!
How do humans deal with it, can they upload a notice to all the other human driven cars that the road is blocked?
Also if the wrongly parked car is self driving, the blocked car can just ask for permission to pass and the wrongly parked car makes way and then comes back to the original position... that is if the wrongly parked car didn't send a preemptive signal to all the other cars in the area that it would be blocking traffic.
I'd be fine with a car that was automated only on highways. I could drive manually on local routes, and then for big journeys I would just head on the highway and turn on autopilot and take a nap, and wake up the next morning in Yellowstone or something.
It seems like standardizing information about just highways would be magnitudes easier than getting every single local route right. Companies could work with the government to standardize traffic and construction pylons/signals to be ideal for detection by robot cars. Traffic, construction, weather, and accident information for highways could all be standardized.
Wow, I never knew you had to have a 3D map of the road you wanted to use and it was so much a hassle, first time I read it after so many articles on the subject, I feel a little cheated.
A friend of mine working for Delphi Automotive (electronics supplier for car makers) recently told me that the google's autonomous vehicle needs exact 3D maps, and at that time I found it hard to believe. Watching Eric Schmidt's youtubes you get the impression like the technology is already there for the car to drive everywhere in the US. Coast to coast distance is 3000 miles as oppossed to 700.000 the car has already driven.
Same as I do. So 700.000 miles is 700x1.000, which is like saying that Usain Bolt has run a marathon in less than 4000seconds (a bit more than an hour)... As long as he has run 400 times the 100 meters in about 10 secs, that would be "true."
Well its not that weird actually. When i drive the first time over an unknown route i make quite some mistakes. Because i just dont know the route. I miss signs etc.
If you drive a route very often, you learn about the route, and you learn the appropiate speed etc.
The advantage of a robot car, is that it knows ALL the streets in advance.
It's almost of no consequence. Assuming the regular cars can do the data collection necessary, the first 10 times a road is driven on might need to be manual but from there, enough data about that road has probably been collected to make it autonomous from that point on. It's still an amazing technology.
I don't really buy the crux of the argument much, which basically boils down to the idea that the car will have trouble with unmapped objects. I would be pretty surprised if all these cars didn't come equipped with an ability to phone home and update the central repository of maps with newfound objects. Thus, when encountering a new object the car could slow down, devote some processing power to mapping it, and then cars traveling through the area in the future should get the latest update downloaded and can handle the previously unmapped object automatically.
If you think about it, it's a really fun problem to get to solve, wish I was working on it.
I agree that "never" is a pretty strong word to use in the context of computing. But I don't think what you described is the most interesting or difficult aspect of the problems that remain.
We're talking about a car, not a mobile phone. It's a 3000-pound chunk of metal that moves fast enough to kill anyone who comes into contact with it, and sometimes even those who ride in it. The ability to consult with a remote server would be nice, but the car should perform just as well even when a neighborhood prankster jams the cell & GPS signals.
So the entire approach of relying on a map might be misguided, regardless of whether the map is precompiled or JIT-crowdsourced. It seems that the current generation of autonomous vehicles rely too much on maps and too little on situational awareness. The next generation will need to make a lot of advances on the latter. Ideally, a car should be able to make all millisecond-by-millisecond decisions by itself, offline if necessary, and use the map only as a hint.
And google's approach is not the only one that scientists are exploring at the moment.
Some are relying on more "reactive" strategies see e.g. Professor Alberto Broggi's work with the university of Parma (http://en.wikipedia.org/wiki/Alberto_Broggi). His vehicle doesn't make use of such an heavy map as the one used by Google.
What I liked in this article is that it reveals to the public that google's communication on the topic is really skewed. They try to make people think that the problem of autonomous driving is basically solved while many big challenges remain.
To be fair, most of the hype around the self-driving car don't come from Google, they come from people who are ... let's just say enthusiastic but who lack exposure in the subject matter.
I agree that the argument is pretty weak. Even if it mapping is complicated and hard to scale, I think that cities themselves would go to great lengths to make their roads compatible with self driving cars. Cities could fund the mapping themselves, or could probably be convinced to install sensors at traffic lights that improve sensor reliability on the cars.
I had the same reaction. This article says "there are some hard challenges" and then equates that with "it's impossible."
The cars at this early stage require everything to be meticulously mapped, but I'm sure Google are working hard on making them handle unmapped situations as well. They have a lot of sensors; surely at some point they can start relying on them for unmapped situations.
I don't know how the system is currently implemented by automatic traffic signs recognition is already possible, see for example this video: https://www.youtube.com/watch?v=hU7yHQkg-7U
I think that is another argument against Google's approach. If one can reliably detect traffic signs, why build that database? For traffic signs, one could argue that it helps increase recall; if the car knows the locations where traffic signs were seen earlier, it can decrease detection threshholds for those locations. However, the car still would have to be able to reliably detect new signs (say for a temporary detour) on first sight.
Things are way worse for all kinds of changes to roads. Even if the first car correctly classifies that white spot on the road as a lost paper that it can drive over, what good does that do the next car? The wind may have blown it away or to a different location and into a different shape.
To me, Google's approach seems an attempt to build a model of what the entire world looked like a short while ago, while cars only need a rough model of what it looks like now.
Scaling Google's approach to millions of cameras in million of cars may improve the model and decrease its latency and might make the latency low enough, but I don't see why it would be the best approach.
I always wondered if it was possible to have a dedicated autonomous car lane on the highway, similar to a HOV lane. To get on, you would park your car at a designated area. The computer would start to sync with all the other cars traveling in the lane, and it would take off and just add itself to the caravan of cars. You could go to sleep or do whatever else you wanted. When you reached your destination exit, the car would park itself again and wait for you to resume control of the car for the local streets.
You still have to deal with things mentioned in the article: random objects on the road, rain, sunlight, human driven cars crashing into your lane. But I think that's more manageable then having to deal with humans crossing the street, stop signs and traffic signals, random road changes in the middle of the night.
Something in this vein--or some other variant of autonomous cars driving on (relatively) predictable limited access highways--seems a much more achievable goal for, say, the next decade than general purpose autonomy. As you say, there are still obstacles to overcome to which I'd add legal and regulatory concerns. I sure don't want a car for which I'm liable if it makes a "mistake."
The general purpose robo-taxi will almost certainly happen someday but I'd be more likely to bet on fifty years than ten.
Liability, lawsuits, litigation, lawyers. The future of autonomous cars is not limited by engineering smarts. The future of the industry will come down to how punitive the jury feels when Google Inc. dragged before the courts after a Google vehicle kills people. And these vehicles will, of course, kill people. But I don't see the "they statistically kill LESS people" argument working against "a machine made by Google ran over my children." Someone needs to be sueable, right?
An Uber driver attacked a passenger with a hammer. Another driver ran over some people. Neither of these events have had any appreciable effect on Uber.
T'll be interestin to see how the public reacts when a Google autonomous car kills someone.
And so many people currently use Google services I wouldn't be surprised if people have already died as a result of something Google has done.
In the Uber case you can sue the driver. So when deep-pocketed Google is the driver, I wouldn't be surprised to see people throwing themselves under the vehicles.
(If anyone is about to retort with "webcams!" you're missing my point. Someone needs to be sued. If Google convinces law makers to enact laws to endemify them, then the State will be sued first, then Google.)
Self driving vehicles are a new transport paradigm, potentially. The trajectory of these things is complicated with all sorts of intertwined positive feedbacks between technology, infrastructure, preceding technologies and succeeding technologies.
Roads were built for oxcarts. They followed routes used by donkey trains which followed paths used by walkers who followed pre existing animal runs wherever they could. The width of road cars and train tracks were based on the width of old roads. Shipping containers were designed to fit on trains and trucks. Ships were built to handle shipping containers, as are ports, depots and such. If you want to use something other than a standard shipping container, you probably need to design it to fit in to the shipping container world. It can all be traced back to wild goats making a path from one place to another.
The development os self driving cars as a major mode of transport depends on the development of stuff around it. Infrastructure is probably the big one. Our roads are built for human drivers in standard cars. If when roads start getting features designed specifically for robots, the whole thing could accelerate. I don't just mean physical infrastructure of roads, but maybe the whole ruleset and/or economics of it.
Many old cities have a problem with 10 sq Kms of inner city traffic. Some ban or limit vehicles here. But, if people use robo taxis instead of cars, these inner cities can have different vehicles. Instead of big cars that must deal with trucks and highways perhaps inner cities can be handled by slower, lighter and safer like vehicles like golf carts. This might help solve some safety issues.
"The autonomous Google car may never actually happen."
Directly underneath that, an image captioned: "A Google self-driving car maneuvers through the streets of Washington in 2012."
This article is rubbish. It makes lots of arguments that sound valid, but are actually nowhere near being insurmountable. Take the first one, for instance. Supposedly, the need to map the roads is a huge burden. Well, what if they just design the cars so that when a road is unmapped they need to be manually driven, but after that has been done a certain number of times, can be driven on that road autonomously? Only a tiny proportion of drivers would even encounter that situation, and even new roads would become autonomous-compatible from day one. (or thereabouts)
Overall, the article acknowledges that varying degrees of autonomy are already being built into vehicles. To not come to the conclusion that these will be iterated on and become (more or less) full autonomy is short-sighted.
All the effort is spent for safety. But is it worth it? The fully autonomous approach may never make Google's self-driving car economically viable to reach the same level of human driving safety. Note that humans are actually decent drivers. WHO reports 7.6 road fatalities per billion vehicle km in the US. Now Google's self-driving car has reached 1 million km, and that's under ideal conditions and human monitoring. If it wants proper validation of the safety of autonomous driving, it requires at least several billion miles of driving, which would cost much much more than a lot of low hanging fruits of improvement at driving safety (autobraking, obstacle radar, auto lane keeping etc).
Now that you can see a slow version of self-driving car already out there with 25 mph top speed and impossible to cause serious injury in the first place. This might be the future of self-driving cars, to provide accessibility and enable those who can't drive.
> If it wants proper validation of the safety of autonomous driving, it requires at least several billion miles of driving
No, it doesn't, the same way we don't need decades of testing to build an edifice or centuries of flying to test an airplane. Actually, as every human driver is literally a different person, by your logic we would need driving tests enduring several billion miles.
We can make tests using the worst situations, corner cases and even simulated accidents to see how the driving AI reacts. The problem is hard, but engineering is a finer art than what you imply.
Note that I was comparing with Google's methods of testing self-driving cars: idle roads, good weather, precision maps, human monitoring, no accidents. If Google's miles of safe driving mean anything, if means it has 99% more miles left before a level of validation. If you look at it for the effective miles where actual accident recovery happens, it will be much much shorter.
And yes, UMich/Ford's self-driving car project is making a testing field specifically designed for increased hazardous environment. But that environment is always artificial, and the hard part is to catch the last 1% situations, or the last 0.000001% in order to reach a level of billion miles safe record, because you don't even know what those are.
Autonomous navigation is fancy on all side (Yes I do this research), but all field experts know any security audit can probably reveal a bunch of failure modes because nobody has really worked on making it robust again adversaries.
Road fatalities are probably also that way because of the cars themselves (crash-tested and with airbags and all), emergency infrastructure (ambulances and medical care) and the relatively high number of 'easy' kilometers (highways). The European average is 13.
But shouldn't we also look at injuries, both physically and mentally, and perhaps even material damage? That probably gives a much higher resolution to compare those 1M km's from Google with.
They forget that Google also own a high resolution satellite company and they're working on a drone project - all which could be very useful for real time mapping.
About parking, this is less important for a taxi company.
And the incentives for self driving cars are huge , even as regional taxi companies, so connecting the traffic lights into the net(read only) as backup doesn't seem like a big problem.
The real interesting question is: does it really need "generalized intelligence" , or all the examples the article mentioned could be coded decently enough on a case by case basis, to a point where self driving car is much safer than a car ? We don't yet know the answer, but the current safety of Google's car are promising in that regard.
Instead, it will simply be the autonomous BMW, Volkswagen, or Daimler. Or Tesla. It's not as though Google is the only company in the world working on autonomous driving. They're just the only one the American tech media has noticed.
There is no reason to assume those European companies have solved this problem anymore than Google has. They most likely will come up with the exact same "solution" with the same limitations.
No. I know what the Europeans are actually researching. Google seems usually to think in terms of the knowledge they gain from universal data availability. Driving, to Google, is an outgrowth of mapping.
The Europeans are looking at a scaled approach of driver assistance. There are cases where automation is already making driving safer and easier, on the road right now in production cars. It's a much more holistic and interesting approach, and while it will make use of data if available, it's not locked into it.
So no, they're not coming up with "the solution" - they're coming up with a wide variety of solutions, with much fewer limitations in the aggregate.
I thought the key point of Googles new tiny & slow self-driving car is that it is so slow, that "just stopping" is the adequate solution to all unforeseen issues. This would give the driver enough time to wake up & take over. (Or someone does it remotely). And really, from all issues that might be open, I am sure that traffic lights are a solved problem. How hard would it be to pass a law that requires all (mobile) traffic lights to send out some radio-signal to communicate with self-driving cars? Such an emergency stop signal would also be useful for human-driven cars.
You don't need a map of the whole world/country in order to operate a regional taxi service. Nor do you need intelligence for parking (except in the taxi garage - which can be done by humans if necessary). You just need a smaller map of the region that the taxi will operate in, and the ability to "phone home" when it sees new objects in the map (which can be analyzed by humans / really big computers and added to the map in real time).
The level of new insight this article provides is summed up well by the author quoting a comment on someone else's article[1] about the challenges self driving cars face before the second paragraph is through.
There are other companies working on self-driving cars - BMW, Audi, Tesla, others. The darpa grand challenges in 2005 and 2007 didn't rely on this kind of detailed map data.
Will self-driving cars ever be able to handle all possible driving situations? Probably not. But humans can't handle all driving situations either. A more relevant question is whether self-driving cars will be better than human drivers. It's silly to say that we need strong AI for that, since we've already seen several prototype systems that do better than humans in many situations without strong AI.
The other thing that this conclusion ignores is what the changes to the road system might be. If everyone gets semi-autonomous cars for commuting, and they work fine except for one intersection where people need to pay attention and negotiate it themselves, then there will be pressure to change the intersection. Maybe besides the carpool lane, you might eventually get autonomous lanes. It's a bit chicken-and-egg, but if there is a system that is useful enough in some situations for people to buy, then it will progress from there.