# Most likely form, and progression to, near-future self-driving vehicles built in countries with current road infastructure?

Everyone talks about self driving cars in the future, and I'm sure at some point vehicles driven by humans will be phased out, driving accidents are pretty much the number 1 killer of those below 40, not to mention issues of congestion and traffic.

However, what form will the most plausible non-driven vehicle take. We have the Google car, which tries to drive on our current roads with extensive sensors. As a programmer I know how hard a challenge that would be. An easier task to accomplish would be a network without any human drivers and with sensors built into both the road and vehicles to help them self-drive, made easier because without humans allowed in the network your automated systems can be far simpler and trust to some road navigating computer to handle planning vehicle paths to ensure maximum efficiency without accidents. In fact I would imagine anyone building up an infrastructure from scratch on some new planet would likely go this route, we could almost do this now if we we had the resources first world countries have now and were building up our infrastructure from scratch.

A third option may be simply creating a public transport system which is more efficient to the point that people don't own cars at all, but to do that you would need a much better public transport then we have now, one that doesn't slow people down waiting for constant loading and off loading of new passengers.

However, we aren't building from scratch. Thus the aim of my question. Given a first world country of today, with existing road infrastructure and drivers etc, how will we progress to automated vehicles? Will we have to go the Google car approach? Will an entire new network of smart-roads that only automated vehicles can drive in have to be built? Will some hybrid exist? Basically given real world constraints on economics, first adopter issues (when only 10% have auto-cars you can hardly write rules that penalize non auto-cars for the sake of the auto ones), and politics.

In short this is less about the technology itself, though still quite relevant, but also how real world logistics of implementation may modify the approach and thus ultimate destination of such vehicles.

• Something to keep in mind is that road infrastructure is more expensive than cars - the stats I see show something like ~90 billion/year spent in USA on cars and ~900 billion/year on constricting highways and roads. So, as a rough estimate, replacing every single car would be cheaper than replacing a small part of the total infrastructure. – Peteris Nov 23 '15 at 19:15
• Something I find interesting is that we already have street-legal vehicles that automatically perform some maneuvers. Audi has cars that can park themselves, and Volvo has a package that can automatically warn and then correct certain forms of driver error, including veering out of lane. I think we will see a gradual shift as these features become the norm, rather than a sudden shift from no auto features to fully automated vehicles. Something worth considering. – wwarriner Nov 24 '15 at 1:02
• Remember there are hazards that aren't other vehicles. The system needs to cope if a horse runs onto the motorway, or a crowd of drunken humans, or a tree falls, or all sorts of other potential failure cases. – Tim B Nov 24 '15 at 10:16
• @Peteris Your numbers for relative spending appear significantly off and backwards. I'm getting ~$210 bn for road spending ($54bn/year federally and $153 bn for state/local spending ); vs the US automotive industry making$524 billion (this may include replacement parts, but at 32k/car sold isn't far off) last year. – Dan Neely Nov 24 '15 at 20:17 • @DanNeely with some more looking it does appear that you're right, and my initial searches for stats was misleading. I cannot find any good measurement of sales volume of new cars, as opposed to revenue of US-based companies with global production or manufacturing amounts - car import/export is huge; but yes, if there are ~17m vehicles sold annually then it comes out to half a trillion and public spending on roads is ~200m/year - I had looked at the revenue of the road construction industry that apparently has huge business in other types of construction as well. – Peteris Nov 24 '15 at 20:35 ## 9 Answers # The Problem Self Driving cars need to do several things: 1. Go from point A to B using roads and only roads (no going through houses, yards, etc.) 2. Obey traffic rules on said route (stay in your lane, stop on red, go on green, yield to the right, obey speed limit, etc.) 3. Account for other cars on the road ("auto-cars" and human-driven ones) 4. Avoid all sudden barriers (children/animals in road, fallen trees, etc.) 5. Deal with sudden, unpredictable shifts in road conditions (sudden storms, black ice, etc.) 6. Do all of the above while not exceeding 5g's of acceleration. (Considered the "safe" limit of acceleration/deceleration in car crashes..) 7. Be reasonably economical/safe when compared to human-driven vehicles The solutions you've mentioned are good and bad at solving these. Let's take a look at each approach. # Guiding Road / Infrastructure Approach This is potentially great for 1 and 2. You can easily signal when a car needs to check certain conditions, because the road can simply tell them when to go. This solution is not so great at 4 and 5. You would need to monitor all segments of road (which is a lot to monitor) and then process what is a threat and what isn't a threat and then notify the vehicles on the road. Did I mention many segments of road are desolate segments of highway where this monitoring goes to waste? That's because the American West is not as densely populated as, say, Europe, or the American East Coast. Of course, a smart road could only look within a certain distance of active cars, but then there is the fact that someone could rig their car to not appear as an active car but still drive, or the simple problem of installing all that infrastructure. Additionally, what happens if a dysfunctional car can't connect with the road? We should also mention that some roads get replaced very often. Asphalt is a great material for roads, but it develops pot-holes and needs replacement. Dumb roads are already costly to replace. What about the smart roads? # Fully Automatic Car The fully automatic car has an advantage of all computing happening on-board. To solve problems 4 and 5, the easiest solution is to put the things for dealing with those on the car. The main solution for #4 is to put a LADAR sensor on top of the car and identify things in the road as you go along. If the car identifies things as it goes along, why not identify road signs and other cars? This means that one system solves multiple problems! In fact, this one system can solve 2, 3, 4, and 5. That reduces the price and the economic load on the infrastructure-builders. #1 will need to rely on GPS, which is a very dependable technology now-a-days. This single technology solves too many of these problems to ignore. So I see the google car approach to self-driving cars to be the most feasible. It requires the least amount of change for most parties involved, so if it's economically competitive, it stands a good chance of being the self-driving car of the future. • good points. one side comment though, the infrastructure approach could use closed off 'roads', like subway tunnels, to minimize the danger of outside interference, and potentially only have supporting hardware at intersections, leaving straight roads alone (or perhaps including some easily sensed 'guider' to help cars align, which could be as simple as a special paint that is easily picked up. Doesn't invalidate anything here, but does make the infrastructure option potentially a bit easier. – dsollen Nov 23 '15 at 17:29 • @dsollen and SUPER expensive! At that point, the builders may as well be building public transport infrastructure, such as sub-ways, instead of "self-driving cars" – PipperChip Nov 23 '15 at 17:31 • which is why i went back ad added the public transport option, maybe the eventual future will be a hybrid, you provide the car and plug it into our public transport, or just public transport where the 'cars' are owned by the goverment and you jump on the next available one etc. I imagine long term this will be the best route, closing off access allows much greater speeds and safety. With increased traffic the speed gain will justify the cost eventually, at least for many areas; like inner city. but I do agree that isn't the near-future approach. – dsollen Nov 23 '15 at 17:37 • #1: I think GPS should be used as for an intelligent-human-driven car, as a hint telling the driver what might be a good route, nothing more. The car should still compare GPS advice with reality (don't turn to the left when GPS tells you if the driver of the car, human or computer, cannot actually see a left turn). – gnasher729 Nov 23 '15 at 23:00 Right now, the Google car approach seems the most feasible one. The reasons are primarily economical; every car already has a bunch of electronics installed, you only need to add a couple of extra sensors to make it self-driving. All the rest is in the software, which is horifically expensive to develop, but once development is done, it can be copied and installed basically for free. The alternative of building "smart" roads is less reliable, horribly costly (because you need to install electronics all over the road network, which comes out as a lot more than you would have to attach to the cars) and with really complicated maintenance (a self-driving car can drive itself to a checkup; to fix something on a road, you have no choice but to go there and bring the tools). The nice thing about the machine perception approach is that you don't need to change much about the infrastructure. If you wanted to be fancy, you could maybe go and install special retroreflectors on highways to allow more precise self-location when you're driving really fast or something, but you don't really need that. It has already been demonstrated in research that you don't need a central planner to get near-optimal pathing for groups of drones, since a collaborative peer-to-peer approach can do the trick. So, once there is enough smart cars out there, you could try to figure out a protocol for them to talk to each other (which, for safety's sake, will probably have to be standardized by law). Thus, on roads used prevalently by smart cars, traffic will be smoother, more fluid and as a consequence could be faster as well. Determining how public transport would interact with smart cars is a bit tricky, since these are two rather distinct use-cases and I dare not venture a guess here. As for legislation, however, once self-driving cars go from super expensive prototype to something relatively affordable, you can easily encourage adoption without necessarily penalizing the majority of users. If the smart cars really do turn out to be much safer (and perhaps even fuel-efficient), insurance premiums, which factor heavily into cost of ownership, are likely to plummet. • "a self-driving car can drive itself to a checkup; to fix something on a road, you have no choice but to go there and bring the tools" - clearly what we need is self-driving roads. – immibis Nov 23 '15 at 23:44 • I worked with someone actively researching drone-swarm pathing and spreading optimization in the absence of central planning, and helped with some of the experiments. Absolutely can vouch that the algorithms work very well. – KRyan Nov 24 '15 at 2:28 • @KRyan yeah, so did I, hence why I mentioned it in the first place:) – Mike L. Nov 24 '15 at 7:23 • @KRyan [offtopic] any reference to start reading about such algorithms? – talex Nov 24 '15 at 12:50 • So what should the self-driving car do when a child runs out in front? Will it brake but inevitably hit the child, or deliberately crash into parked vehicles on the other side of the road? Will its decision take into account how many people of what ages and states of health are in the car? Should it? – nigel222 Nov 24 '15 at 19:51 Personally I think that should a car ever get to the point that it can autonomously drive the rural one track roads that I drive daily, it will be a strong AI in need of emancipation! ( Decisions recently taken: drive over a small fallen branch, apply brakes hard for a puddle because nothing behind me and a pedestrian off to the side, and how/ when to overtake the riders without scaring their horses) I think a likely development will be "slaveways": major roads or reserved lanes in which only vehicles with a legally mandated control system are permitted. Once you're on one the car and roadside automation do all the work and you can sleep or read a paper. Traffic density could safely be far higher because each car "knows" instantly what the ones in front are about to do. Or the cars might even couple up into trains. Elsewhere, you drive and the car monitors progress. If it detects you are headed for a collision it will brake or possibly swerve before you can. ISTR collision avoidance is already in some production cars. • My thoughts exactly! What you need to drive cars are humans, minus a big chunk of unpredictability when you get them in groups, so smart AI perhaps. – Kzqai Nov 23 '15 at 19:26 • The Interstate Highway System in the US would be a prime candidate for the first of the 'slaveways', since limited access makes the job of the AI simpler. It could even be a bidirectional requirements: only auto-enabled cars on the interstate, but they can only self-drive there (i.e., manual control on surface streets). – Ghillie Dhu Nov 24 '15 at 0:04 ## Public transport We haven't even finished automating our existing public transport on dedicated travel ways yet. There are automated subway trains, so they're further along than Google is. But they haven't finished converting. Many trains still have human personnel. New public transportation is expensive. For example, the California high speed rail system is currently projected to cost \68.4 billion: https://en.wikipedia.org/wiki/California_High-Speed_Rail

And that's a replacement for just one road. Note that the federal government only spends \$80 billion on transportation. So if it bought one high speed rail segment a year, it would have hardly any money left for maintenance of the existing roads, already underfunded. If we include local and state government spending on roads and water as well, we can get up to \$416 billion. That's still only enough for six segments even if we drop all spending on water, which is rather important. And that's only 3120 miles of travel way (six 520 mile segments). Note that the US currently has 2.6 million miles of paved roads. Even assuming that a million miles of that are redundant, we'd still need five hundred years to replace the existing road system.

Unless you are willing to increase spending on transportation construction by a lot, this seems totally infeasible. Who wants to wait five hundred years for automated transportation?

It's possible that high speed rail is more expensive to build than slower rail systems. Ohio's proposed more modest speed rail system was only expected to cost \$3.5 million per mile, not the \$140 million per mile of the California system. That's better. If we only double spending for twelve and a half years, we could replace the system. Of course, that system would also be more complex. Rail lines would cross each other. People would have to switch trains relatively frequently, almost as frequently as we switch roads. That complexity would bump up the cost again.

Also remember that we have to maintain the existing roads while we build the new system and maintain the new system going forward. And what do we do with the obsolescent roads? It will cost money to block them off or tear them up as well.

And note that this requires everyone to double the money spent on transportation. Sort of feasible for the federal government but difficult for local governments. A more reasonable alternative would have local governments maintain existing spending while the federal and state governments each put up enough for a new segment a year. That extends the process out another twenty-five years.

That assumes that the Ohio projections were correct. Note that the California numbers increased by a lot from initial projections. If we quintuple the Ohio projections (which is the kind of cost increases that were in the California project), we get amounts more like what California is showing. And that assumes that the current California projections aren't still underestimating the costs. Even only doubling the costs, that takes us out seventy-five years.

Automated roads that manage traffic are better at avoiding car to car accidents, but they are worse at everything else. For example, what if a child decides to run across the highway? The road's automation can't control the child. So cars running on that road need to have just as many sensors and ability for independent reactions as the Google car. Otherwise, they are less safe.

I think that automated roads are likely to occur at the end of the process rather than at the beginning. When many cars are automated, they'll start getting their own lanes like high-occupancy vehicles do now. They'll join in linked convoys so that most can shut off their motors and coast. But this requires cars to be automated first.

Another issue is that the automated roads would only work with automated cars. If you put regular cars on them then the automated cars have to handle non-automated traffic. So they require automated cars to function. And how do you get to the automated road? You'd have to take legacy roads to get there. So the automated cars could not require the automated road, which just gets us back to the Google car.

The Google approach is the only one that works in the near term, that allows automated and non-automated cars to share the road. That allows automated cars to make it all the way from the initial location to the destination without switching vehicles. And most people won't buy if their car won't go everywhere they need to go.

The Google approach is clearly the most difficult to program, but tests show it to be feasible. If it's feasible, then it's easier in every other way.

There is no evidence that people are interested in increased public transport. It's inflexible, inconvenient, and expensive. Most people would prefer to take their cars with them rather than switch to public transport that runs on its schedule, not theirs. Also, public transport only works if you're going to a place covered by public transport. Many people aren't. You can drive to public transport, but you can't drive your car after you leave it.

The Google car approach is the most likely because it doesn't rely on other changes. It scales anywhere from one to a billion cars. The other approaches require a certain level of usage to function. It can also be implemented on an individual basis. A single individual can choose which car to drive. A single individual cannot choose to implement public transportation (unless really rich).

Without the Google car, we'll never get to the point of being able to use automated roadways. We also won't be able to make taxis cheap enough to make public transport practical in general, not just when traveling from and to urban locations. For most people, the Google car is the only method that they could use.

• Two notes: 1) Building a earthquake proof safety rail is really challenging and increases costs a lot. 2) I can go with public transport, it's by far the best cost benefit wise but note a real total public transport needs to put to use all ways fo locomotion avaiable, trains, subways, bus, minivans, public self driven cars, bikes and your legs – jean Nov 23 '15 at 18:54

One of the main selling points of self-driving cars is that you don't need to buy one. Instead they will operate as taxis. The most expensive part of the taxi is the driver's time; once the car is self-driving it will be much cheaper.

(Fun fact: the taxi service Uber is actively working on self-driving cars.)

Currently I think everyone is designing self-driving cars with the intent that they share the road with normal cars. It's true that we could gain some efficiency by having self-driving-car-only roads, but during the transition period we will need cars that can share the road anyway. If we get multiple manufacturers making self-driving cars, it will be a giant headache getting them all to intercommunicate, so that's probably a long way off.

Here in Germany (particularly in Stuttgart) we have "car to go" cars. They are electric drive Smarts. There is nothing autonomous about them, but you can rent them one-way when you need them, and use a smartphone app to locate and book the nearest one.

Obviously Google and Tesla are working on self-driving cars, and they are not the only ones.

While owners of cars often have the desire to drive the cars themselves, people who rent cars sporadically are typically people who use public transport a lot and thus find it more natural that someone (or something) else is doing the driving.

The way i could imagine self-driving cars to become more popular would be a combination of the above: Self-driving one-way rental cars in cities. It would make a lot of sense in every way. The price, design and inspection intervals are irrelevant for their users. The ability for the car to drive autonomously would help getting them to the garage for maintenance, getting them to the user once he needs them, re-arranging them if they should cluster in some place, and even allow using them for the ride home after a party, if some (or some more) alcohol was involved.

Over time, people will get more and more used to them. The technology will become more robust, and due to installed numbers cheaper, plus people will have more trust in a technology that has proven itself over some time. Eventually, more and more cars will come with these features, until it will become near inconceivable not to have the self-driving systems integrated in a new car.

One note: while i think that this way a very large percentage of all cars could become autonomous, i doubt this will reach 100% any time soon, since there will always be people who either don't trust the technology or keep their old cars simply because they like them, or can't or don't want to pay for a new car, or because they need the ability to handle things for themselves (think bodyguards or military, where the chance of someone hijacking the system is an unacceptable risk).

• a good point, autonomous taxis may be the first real use. I imagine eventually once autonomous vehicles reach a certain critical mass the government will mandate only them being sold due to the extra safety advantage, though obviously that will be a bit after the first adapters. – dsollen Nov 24 '15 at 15:17

The future is already here, as it were.

There are some concept self-driving cars which use cameras to identify obstacles, lines on the pavement, etc. Sensors are nice to have, but ultimately not needed.

• "Sensors are nice to have, but ultimately not needed" ... until a child darts into the road to fetch their toy in front of your car. A passenger may not get to the breaks in time, and certainly would enjoy their cars respecting human life. – PipperChip Nov 23 '15 at 17:22
• Even the most advanced version of Google car they don't trust driving without a backup driver. And it's absurdly expensive to support all those sensors. We are still quite a ways away from a viable self driving vehicle that the average man can afford. – dsollen Nov 23 '15 at 17:24
• camera as sensors....scary lol. We are...not that great at processing video feeds and making sense of them; in fact were pretty horrible at it right now, though we have made good strides recently were still a ways away. Simpler sensors are much easier to program to, like sono to decide how far away a large object is. Lots of simple sensors would add a bit of expense over a camera, but would make programing much easier, and as a side effect increase safety since your less likely to have 'bugs'. it doesn't add that much cost anyways. – dsollen Nov 23 '15 at 17:33
• @dsollen cameras are super cheap (starting at ~10\$) and, speaking as someone who works in visual computing, we're halfway decent at detecting stuff in video feeds. Certainly if you have multiple cameras, you can detect obstacles (ie. something is there but I don't know what) quickly and easily. Convolutional Neural Networks are pretty good at detecting specific objects, but have yet to be made robust to practical situations and sometimes misbehave in funny ways. – Mike L. Nov 23 '15 at 17:42
• About the "human backup": A human cannot possibly take over quickly. It's the whole point that the human can sit back and relax (most over the time). If kids jump into the road then the car needs to handle that by itself. Where the "human backup" works is when the car encounters a situation that it cannot handle, comes to a stop, and let's the human driver take over (possible after separating from the second passenger on the back seat :-) Say the self-driving car can't handle driving in deep snow, that's when the human has to take over. – gnasher729 Nov 23 '15 at 23:05

Road sensors seem highly expensive, far more cost effective could be a special paint which the car's computer can easily identify, with plenty of built in redundancy in case of wear, poor visibility due to weather, etc. Instructions to the car could even be encoded into the colour or patterns of the paint. Signs posted on the side of the road containing the equivalent of QR codes could also be used. No need for fancy expensive electronic sensors.

But I agree with others and think it'll be more important for everything to be self-contained within the car, and over-ridable by the driver (or actually passenger I suppose), the additional reason to what others have said being you may want to drive somewhere where the infrastructure has not been laid down.

For example, sometimes big once-off events/festivals, rather than being forced to be near an established car park, simply pay a local farmer for the use of one of their fields for patrons to park their cars. Right now it just needs to be an open, flat surface. How would a car which relies on road sensors park here?

## Test it in fishbowl territories

By that I mean, find areas with modern roads that are isolated from the rest of the world by physical and/or political boundaries. The biggest problem you'd have trying to implement such a system in the US or Europe is the sheer number and length of connected roads over such large areas. What you want is a place with a lot of money but very little area or connection to surrounding territories.

A good example would be something like Hawaii. It's a relatively rich state with a small area, and not connected by roads to anywhere. It would be a relatively simple task there to convert all roads and vehicles to be sensor-equipped. Other examples might be Japan, Taiwan, Ireland, Iceland, and any other island nations. Examples of non-island nations might be places like Israel, which despite having land borders with several other countries has very little cross-border traffic for political reasons. South Korea is in a similar situation, as its only land border is not exactly a friendly one.

In fact, there are some similar projects being considered for places like these. A system for battery-powered cars has been proposed, where the batteries would be rented rather than owned, and could be replaced with a fresh battery on the fly at battery stations, which would eventually take the place of gas/petrol stations. The challenges to infrastructure development would be similar, so similar strategies for test-markets would apply.

Once you have established systems up and running in a few prominent isolated locations, you now have proof of concept, and you can point to the likely dramatically reduced instances of vehicle fatalities in those places due to your system, which will cause interest in larger nations to increase. I still think places like the US and Europe will be fairly late adopters, just due to the sheer scale of such a project. You might see it cropping up in places like Switzerland first, or possibly in major cities like New York, San Francisco, L.A., Chicago, London, Paris, Berlin, etc. long before it takes hold in the surrounding areas. It's possible you'd even have "hybrid" vehicles, which are autonomous in major cities where the road sensors have been installed, but still under human-control everywhere else.