The truth about self-driving cars

Regular readers will know I’m rather skeptical about the prospect of self-driving cars (1, 2, 3) and so I listened with interest about what somebody in charge of a car manufacture had to say about them.

He first listed the five degrees of autonomy, with Level 5 – the highest – allowing the human occupant to remove his hands and switch his brain off any time, anywhere. This is what most people think of when they talk about self-driving cars, the ability to go to sleep in the back or be blind drunk leaving the car to take care of everything. At the moment, production cars are fitted with Level 1, which is basically driver assistance. Level 2, which allows the driver to drop their concentration a little, is being introduced slowly.

He then talked about five technical areas which will need to be tackled in order to have Level 5 autonomous cars. I’ll take them each in turn.

1. Computing

Modern cars currently have around 50 black boxes carrying out various functions. In a fully autonomous car, they will likely have a single computer split 5 ways, with the parts carrying out the safety-critical functions kept well separate from the bits that run the entertainment system. Raw computing power is unlikely to be an obstacle to the development of autonomous vehicles.

2. Antennae and Sensors

The number and variety of sensors and antennae an autonomous vehicle will need is mind-boggling, particularly if redundancy is considered and 2-out-of-3 voting required to avoid spurious trips. The antenna on the Google car can be seen in the picture below:

A fully-autonomous car would need about 5 of these, mostly for communication outside the vehicle. It would need multiple 5G connections as well being able to connect via wi-fi and satellite. Sensors will include radar and infra-red cameras, which must be kept clear of dust, dirt, and rain.

3. Decision Making

Here’s where it starts to get complicated. What does the car do with all this information it’s receiving? The software is going to have to come pre-programmed with every situation the car can conceivably encounter so it knows what it’s looking at. Even if we charitably assume self-learning AI will be fitted to the cars, automobile accidents are often such that the occupants, be they human or computer, don’t get a second chance. The sheer size of this task in achieving Level 5 autonomy for cars is unprecedented.

4. GPS Mapping

GPS for civilian use is accurate to around 3-15m, although considerably better when the US military is lobbing missiles through windows and cave entrances. Level 5 autonomous vehicles will need GPS mapping to be accurate to within centimeters. If the car comes with an incredibly accurate GPS map installed in its brain, what happens when the map changes? A new road could be easily updated, but roadworks? Will we rely on the South Pembrokeshire District Council to inform whoever makes the maps in an accurate and timely manner every time they dig up the street?

5. Control and Action

Once a car has figured out where it is and what’s in front of it, what action does it then take? Does it jam on the brakes, swerve, or carry on? Software that could handle this in a normal street environment is not even on the horizon, and probably won’t be for another twenty years at least.

He emphasised that these 5 areas only cover what is required in the car; the infrastructure required to support autonomous cars was an equally gargantuan technological challenge which national or city governments will have to deliver.

Our visitor compared the challenge of Level 5 autonomous cars to landing on the moon, only without the single, dedicated organisation driving it. He didn’t say whether he thought we could replicate the Apollo 11 mission today, but my guess is we wouldn’t stand a chance. For a start, there is a worrying lack of diversity in the picture below:

I asked him whether he thought, as I do, this is all just a pipe-dream and we might never see Level 5 autonomous vehicles. He replied that, in his opinion, the technology will advance while there is an obvious benefit for the additional cost, as was clearly the case for ABS brakes and traction control. So it could well be that we get to around Level 4 autonomy before the costs and effort to reach Level 5 outweigh any benefit.

One interesting thing he said was that the most obvious place to use autonomous vehicles was on motorways, where the environment is much more strictly controlled than on other roads. The trouble is, only around 3-5% of road deaths in Britain occur on motorways, with the bulk taking place in urban areas or on rural roads. This is because on motorways the relative speeds of the cars isn’t too dissimilar, so in a crash cars just tend to get bounced around a bit while all heading in the same general direction. By contrast, accidents on country roads tend to involve cars converging at speed, hitting stationary objects, or leaving the road altogether. Therefore, the easiest and most obvious place to have autonomous cars will not save many lives, which kneecaps one of the main arguments of their proponents.

He also mentioned the legal aspects of autonomous cars. Currently drivers are responsible for accidents, and individual drivers insured. With autonomous cars, it will be the manufacturers which will be responsible, and this will drastically change the legal and insurance landscape in any country which adopts them. He didn’t put this forward as a reason autonomous cars won’t happen, he just mentioned it as another thing to consider. Regarding the technological challenges, he didn’t think there was any chance Level 5 autonomous vehicles will be possible for at least twenty years. My guess is it’ll be a lot longer than that.

Liked it? Take a second to support Tim Newman on Patreon!
Share

62 thoughts on “The truth about self-driving cars

  1. My understanding of the marketing idea behind autonomous cars is as follows:

    Say you have a neighborhood with 10,000 people. Each has 1 car, that makes 10,000 cars, but 90% of the time these cars are idle. A company will offer ‘transport services’ using autonomous cars. The company will have 1000 cars which will be shared by the 10K inhabitants. You will summon a car using an app; an available one will be allocated to you and drive autonomously to your house. You will only wait a few minutes. You will drive to your location, relinquish the car, which will then drive itself to the next customer.

    The idea is that you will save a lot of money, because you don’t really need 10K cars for 10K people, and the company will be hugely profitable.

    My own opinion is that this will be a disastrous failure for the company, and that if somehow they get it to work, it will be a disastrous failure for the population.

    I think it will fail because of the tragedy of the commons, and because most people need a car at roughly the same time, not spread out throughout the day. I believe that large companies will lobby to have legislation passed to make it profitable nonetheless, as happens with many green technologies. They will extol the number of lives saved, they will magnify the environmental damage of gas-burning cars, they will lobby to pass punitive taxes on private vehicles, etc.

    When it still proves unprofitable, they will reduce the quality of service. People will wait longer, drive slower (hey we have to pick up someone else), and receive dirty cars. By this time, many people will have given up their private vehicles – and buying new ones will be more expensive (see above), so the company will effectively be a monopoly, or part of a cartel of similar companies. Private individuals will be reduced to serfs, and only the nobility will ride around on horses – sorry, I mean in private limos.

  2. Say you have a neighborhood with 10,000 people. Each has 1 car, that makes 10,000 cars, but 90% of the time these cars are idle. A company will offer ‘transport services’ using autonomous cars.

    Yes, our visitor went into this in quite some detail and described the business model. But it assumes the technological challenges can be overcome.

  3. Software reliability

    rinse and repeat

    and umm…. as you point out >> real life

    On a “track” – like a mine, quarry, warehouse say – when the environment can be controlled then fine.

  4. The trouble with the intermediate steps is that they presumably tend to lull the driver into a false sense of security, rather like parking sensors. You start relying on them, drop your guard and then bang something goes tits up and you hit a litter bin or worse, a 6-month old baby.

    A self driving car that can’t drive home from the pub with you insensible inside, is totally pointless.

    Give me a 10 year old RangeRover any day.

  5. Are any of these companies thinking about security? And not in a “can our car be hacked?” kind of way. If you have a self driving car, that won’t let you crash, it becomes very easy to immobilise a vehicle – one person in front, one behind. Then the occupants are at the mercy of the nefarious types.
    No thanks. I’d rather be able to escape.

  6. The most important claim for autonomous cars seems to be that, come their glorious availability, they will kill many less road users than do human drivers. However, there is something that worries me about this. USA traffic deaths per unit distance driven are approximately twice those of the UK. From Wikipedia (2013 figures; see https://en.wikipedia.org/wiki/List_of_countries_by_traffic-related_death_rate ) the USA has twice the road fatalities per unit distance driven than does the UK. This is 7.1 deaths per billion vehicle-km for the USA and 3.6 deaths per billion vehicle-km driven for the UK. On top of that, whilst it is obvious that reduction in First World road deaths is very important to us here, there are many countries with much worse road death rates than the USA.

    Probably there are underlying reasons, difficult to overcome, for some part of the higher rate in the USA than in the UK. But surely getting much of the way towards halving the USA road death rate (and the rest) is something that could be done immediately – not something for which we have to wait at least decades (perhaps for ever) for new technology. Or are governments not really serious about why they want autonomous vehicles – just pandering to a rent seeking industry?

    As a subsidiary point, do vehicles with Level 1 assistance have a usefully lower serious accident rate / death rate (that can be attributed largely to that assistance, and with some reasonable confidence)?

    Best regards

  7. As a software kind of guy, if the self driving car people are relying on ultra accurate GPS maps I can tell you right now they will never succeed. As you say there’s the problem of the local council keeping everyone up to date regarding roads and their navigability. It simply isn’t going to be reliable enough and it is an obvious single point of failure. A Bond villain (or a terrorist) will see that hacking the map/GPS will cause the required confusion/destruction required for his evil plan and it is quite clear to me that no matter what is attempted, if you can update the map then someone can hack the update process and feed in bad data (if you don’t update the map you have the whole “roads not existing” problem)

    However, that doesn’t mean self-driving cars are impossible or even impossible in a decade IMHO.

    Humans drive by a combination of knowing where to go (maps, GPS, road signs) and examination of the local environment (e.g. following diversion signs when a road is closed). An autonomous vehicle must be able to examine the environment and adapt to local conditions to avoid hitting things and there’s no reason, that I can see, why that examination should not also apply to figuring out the road layout and relevant signage.

    I don’t say it is easy, but I can think of a number of ways that existing technologies (such as Automatic Number Plate Recognition (ANPR) ) could be adapted to reading road signs and then rerouting based on following the signs and/or other indications of where to go. Some of that might be as simple as following the vehicle in front, though of course the FIRST vehicle needs to be a bit smarter.

  8. An autonomous vehicle must be able to examine the environment and adapt to local conditions to avoid hitting things

    Well, yes. See Numbers 3 and 5.

  9. It isn’t binary.

    There will be quasi-democracies such as Singapore that will legislate the relevant infrastructure changes, for example.

    Perhaps there will be a toll motorway that encourages or incentivises level 3 or 4 technology cars. That’s quite an attractive value proposition if one were contemplating a 7 hour drive, for example; drive down the on ramp and hit autopilot.

    There’s lots of risks and impediments but technology doesn’t tend to get implemented in huge leaps, but incrementally.

    As for Jonathan’s concern about big business and government colluding to screw us over; I reckon they’ve got far more options available than this one…. and they will!

  10. @Chernyy_Drakon: “Are any of these companies thinking about security? ”

    Almost certainly not enough. And that includes making it secure from hackers as well as your example. There are way too many ways that a self-navigating vehicle could be attacked. Since automated logistics are a major market for self-driving vehicles I predict there will be mass hijackings and an arms race between crooks and vehicle owners/makers once people try to deploy them in any country that has criminals (i.e. every single one of them)

  11. The “autonomous vehicles as a means to rent a travel pod” has been around for at least twenty years (it was being proposed at an INCOSE conference back in 1998 to my certain knowledge) and Jonathan’s concerns were raised then and haven’t been addressed clearly since.

    I look at the examples of the bicycle rental schemes that have sprung up in various cities, and the success (or lack thereof) of them; when even the Guardian is despairingly writing about the abject failures of “dockless bike rentals” in multiple cities, the auguries for a more complex, more expensive and more vulnerable system are poor.

    And just from a practical point of view, how does a “transport pod” work out that its previous customer, ferried home from a serious night out, used it as both vomitorium and lavatory; has it been hosed out and disinfected, and has it had a chance to dry out, by the time it arrives for your morning commute?

    It’s a lovely theory for a science-fiction future where everyone wears silver togas, treats each other and public property with due respect, and Obeys The Computer (because The Computer is YOUR friend!) but it’s not a good fit with observed human nature.

  12. Also, self-driving vehicles have a number of potential benefits. Not causing road accidents is only one of them.

    I could imagine intermediate solutions where there are dedicated motorway sections that are only open to self-driving vehicles (see how the US has HOV lanes as an example). If the goal is (and it will be for logistics) to not have a driver per vehicle for most of the journey, then a completely acceptable interim solution is one where you have lorries self-driving only on the long-distance intercity parts of their journeys using these dedicated lanes/carriageways. The first and last few miles have a driver who gets on/off the vehicle at something like the toll booths you have to get on/off toll motorways*. There’s no reason why passenger vehicles could not use the same infrastructure. The driver can nap/work/whatever during the long-distance part and only needs to actually drive in the first and last few miles.

    *yes there’s the question of what the starting driver does next? wait for a returning vehicle, take public transport (or uber) back and likewise how the finishing driver gets to the vehicle to drive it the last few miles as well as what happens when that driver is delayed (where does the vehicle wait?). But these are all fixable issues and there’s a lot of money to be saved if they are solved so I expect they will be.

  13. If reducing deaths is the key driver then the quickest and cheapest way is to remove the driver’s airbag and seat-belt and install a 6 inch spike in the middle of the steering wheel.

    More seriously, it was either Freakonomics or Planet Money that did a series on autonomous vehicles and they interviewed a lot of the key players. The consensus seemed to be that true autonomy ie a car with no controls for everyday use, was at least 30 years away and not getting closer.

  14. Level 2, which allows the driver to drop their concentration a little, is being introduced slowly.

    Expect a slow but inexorable increase in pedestrian, cyclist and even driver deaths then. The tech will never keep up with the stupidity and laziness of human drivers.

  15. used it as both vomitorium

    Just to be that person for a moment, a vomitorium was not where people vomited, it was the room by the exit; ie, it was where the building vomited (its occupants out onto the street), now there the people did.

    I recommend:

    https://rodneybrooks.com/edge-cases-for-self-driving-cars/

    for more on why self-driving cars are a long way away if they are even possible at all.

  16. Is the self driving car a bit like nuclear fusion – its always 20-30 years away from being a viable proposition?

  17. Robbo–You must be joking. What grown-up wants to drive around in that POS. You might as well save money by re-using old Dodg’em cars. The crapwagons in the film make the Sinclair C5 look like a sports car.

    As for self-driving cars–well, in a truly free and free-market society ( such as we don’t have) they would likely have evolved over the next 100-150 years. And –in a free world–they would be a boon. Go somewhere and send the car to pick up your Missus or stop the car, get out and walk the dog home while sending the car on ahead by itself. All good stuff.

    That is not what is happening with all the contemporary autonomous cars bullshit. Behind this is the political and bureaucratic scum who, having viewed “Total Recall” once too often ( “You’re in a JohnnyCab!”) see the possibility of turning a private transport system into a state-controlled public one. They can steal even more from you and control where and when you go. That is what is behind the hype and bullshit.

    Luckily the AI is light years away and this is one of their tinpot schemes that will backfire and fail badly. However, first there will likely have to be some crashes so spectacular that the CHiPs show would be envious. Before the scum accept the puncturing of their control fantasises.

    They have already managed to kill 3 people and they are nowhere near anything yet.

  18. “Say you have a neighborhood with 10,000 people. Each has 1 car, that makes 10,000 cars, but 90% of the time these cars are idle. A company will offer ‘transport services’ using autonomous cars. ”

    Yawn. You’ve just invented the mini-cab. Or to give it a modern name: Uber.

    Shortly, people will talk about self-driving cars having to have dedicated routes (to keep away the animals and children who will not be 5G enabled) and they will then invent the railroad.

  19. Shortly, people will talk about self-driving cars having to have dedicated routes (to keep away the animals and children who will not be 5G enabled) and they will then invent the railroad.

    https://rodneybrooks.com/bothersome-bystanders-and-self-driving-cars/

    ‘Drive.AI founder Andrew Ng, a former Baidu executive and one of the industry’s most prominent boosters, argues the problem is less about building a perfect driving system than training bystanders to anticipate self-driving behavior’

    So if you get hit by a self-driving car, it’s your own fault for not behaving yourself in the way the robots expect.

  20. I have no particular insight into the technology involved but I understand that Google are having some success in Phoenix rolling out a self-driving taxi service pretty soon. There is a lot of criticism that Phoenix has very predictable roads and weather, but then you need to start somewhere.

    I’ve also read that the difficulty isn’t with the first 95% of the tasks required to get a self-driving car up and running, but with the last 5% which take a the vast majority of the effort. It’s why a lot of companies have self driving car projects now and superficially seem to work but are finding it hard to actually get anything to market.

    I’m not really sure about a lot of the assertions in your post made by the car manufacturer exec. My opinion is that Google et al are not dumb and haven’t thought “oh crap we didn’t realise we wouldn’t be able to rely on millimeter scale mapping of an environment and roadworks could happen without the council telling us!”. I think a lot of the car manufacturers have outsourced their research in this area to companies like Delphi so perhaps the exec isn’t really up to speed with the nitty gritty, or has misunderstood various briefings they’ve been given. I also acknowlege there is tonnes of hype. But personally I’d prefer to wait and see what happens in the next couple of years before saying it could never happen.

  21. Re the car driving you home from the pub, there are jurisdictions in the US where it illegal to be drunk asleep in the back of your parked car. So there’s another stumbling block.

  22. “Re the car driving you home from the pub, there are jurisdictions in the US where it illegal to be drunk asleep in the back of your parked car. So there’s another stumbling block”

    Program the car to circle the block if you get home but don’t wake up to go inside.

  23. The thing with man on the moon is that everyone knew it was broadly, theoretically possible for decades. Build a rocket big enough etc.

    It’s always the example people use: if we can put a man on the moon, why can’t we do … But those other things are often in the realm that we don’t have even a broad solution.

    Self driving cars are in the latter. And the only reason there’s stories everywhere is corporate PR and it’s an easy story for journalists.

    The real changes in transport will be boring and incremental. Sometimes it isn’t even the thing. I started using buses more because Stagecoach did things like e-ticketing, timetables and live bus times. If someone comes up with secure bike parking, we’ll improve that.

  24. Tim writes, probably quoting his visiting lecturer: “Raw computing power is unlikely to be an obstacle to the development of autonomous vehicles.” I would like more detail on this.

    For example, my understanding of AI computers is that the ‘more intelligent ones’ use HPCs, with top-of-the-range ones (providing PetaFLOPS) currently consuming a few megawatts (and associated cooling). Is this sorts of computing power not likely to be necessary for Level 5 autonomous vehicles? If so, where will the computers sit? In-vehicle: how powered and cooled? Or with mobile communications: what are the required data rates from all those sensors with many many vehicles per square kilometre (wireless and subsequently wired/fibred).

    These questions need sensible consideration, on the eventual practicality (of the necessary infrastructure) and on likely timescales.

    There is no point in planning to escape to another habitable planet to avoid the next big asteroid: unless you have already got the warp drive. Better to send Bruce Willis with a bomb.

    And for driving around, maybe it’s better to plan to keep relying at least somewhat on those pesky 25W squishy brains! Even if cyber-you can beat all of them at Chess.

    Best regards

  25. dotdavid,

    “I’ve also read that the difficulty isn’t with the first 95% of the tasks required to get a self-driving car up and running, but with the last 5% which take a the vast majority of the effort.”

    This is always the same with software and why this won’t work.

    For instance, I used to look after a billing system for 1.2 million customers. It printed the bills and went through a mailing machine. For about 1.19 million customers. We then had these cases like blind customers, customers abroad, customers with too many pages for the envelope. Human intervention was required. Someone looked up the postage to Brazil or sent the bill for braille encoding, or took the bill and manually put it in an envelope.

    Every computer system of any complexity has humans for these exceptions. Sometimes that’s about cost, but sometimes it’s because a computer can’t do that task. Checking documents for money laundering laws is a human task. Image recognition just isn’t good enough. We often design processes for businesses with these human gates and sometimes later we remove them. But you almost never get eradication.

    Right now they can do a lot of straight line, good weather stuff, but can they recognise that cars are driving around a good road? Or identify a cyclist?

  26. Tim writes, probably quoting his visiting lecturer:

    No, that was me.

    If so, where will the computers sit? In-vehicle: how powered and cooled?

    Power generation in a car with a petrol or diesel engine isn’t a problem, nor is cooling. With an electric car it’s more tricky, but if the power problem is solved for propulsion it won’t be a problem for running a computer either.

  27. I’ve also read that the difficulty isn’t with the first 95% of the tasks required to get a self-driving car up and running, but with the last 5% which take a the vast majority of the effort.”

    That was exactly his point here:

    it could well be that we get to around Level 4 autonomy before the costs and effort to reach Level 5 outweigh any benefit.

  28. @Tim I actually mean to even get to level 4, or even 3. As Bloke on M4 says it’s a common problem in software, edge cases always take the vast majority of the effort, and designing autonomous vehicles has a lot of edge cases.

  29. “3. Decision Making”

    This one is a no brainer for the “not yet ready camp”, never mind the human/computer interface when you are supposed to grab the wheel in the nick of time to avert disaster, just stick to the limitations of computers.

    Autonomous vehicles work very well and have been proven to work well in the “one to many situation “ in large scale commercial controlled situation where one worker at a central base now replaces does the work that was previously done by many workers at a remote location with very high costs. This works well and has been proven to do so in certain Australian mineral hauling truck case.

    And that is where it ends for now.

    The notion that algorithms are cleverer than humans has been around since computers were first invented, but it has always proven to be false. Algorithms still are and always will be inferior to a human, safety is psychological. Sure, a computer can beat a human in a set environment in a fixed situation, but we humans win this argument every time and always will over a computer in anything else other than that, particularly the very risky and variable situation of driving a very heavy hunk of steel with huge inertia in the general direction of others huge hunks of steel with huge inertia with many humans getting in the way.

    Computers cannot deal with snow, rain, wind, hail, tunnels, road signs, body languages, punk rockers, tolls, incorrect road signs, faulty traffic signals, drink drivers, Volvo drivers, graffiti, misleading signs, blood, fire, crowds, other drivers, bin men, young drivers, old drivers, handicapped drivers, drivers from other countries, bad drivers and fuckwits.

    So yes get with the the program and buy lithium miner shares for EV’s, but other than investing in some mining companies that are successfully cutting costs with “the one to many rule” I can’t quite see a quid in this autonomous punt yet.

    …………………………………………………………………………………………………

    The Big Problem With Self-Driving Cars Is People
    And we’ll go out of our way to make the problem worse

    https://spectrum.ieee.org/transportation/self-driving/the-big-problem-with-selfdriving-cars-is-people


  30. Tim the Coder on October 22, 2018 at 12:12 pm said:

    “Say you have a neighborhood with 10,000 people. Each has 1 car, that makes 10,000 cars, but 90% of the time these cars are idle. A company will offer ‘transport services’ using autonomous cars. ”

    Yawn. You’ve just invented the mini-cab. Or to give it a modern name: Uber.

    Zipcar also solves this problem. The step to self-driving adds very little value to this for a significant extra cost.

  31. In a mine if you walk in front of a truck it is your fault – approximately, still lots of paperwork.

    In a city the “choose who to kill” problem does seem very tricky and not amenable to a technological solution.

    Motorways does seem the most compelling opportunity. I have no idea if there is any analysis of how much more effective capacity you get from smoother driving / tighter packing. How much demand might change (1 hour motorway driving is about as painful as 2 hours where the computer drives and I can read – economics predicts people will put up with worse traffic jams if they are less painful). As to making money from the push for level 5 – chip makers and coders with the right cv seem to be doing very well.

  32. Shortly, people will talk about self-driving cars having to have dedicated routes (to keep away the animals and children who will not be 5G enabled) and they will then invent the railroad.

    Heh…

    Daimler Trucks is working on this a lot. The on-board computers don’t have to be the size of EV batteries and fill your trunk (story, boot). They can also be split up throughout the car. His comment that it would be one combined unit, confused me, though. The other issue is data storage. A giant freight truck (sorry, lorrie) can spit produce 20TB of data per day. The vehicles are connected to the cloud in order to store or use the data. This is a huge amount of data, needless to say. I imagine the computing storage companies are going to do very well in the coming years. Same with cloud vendors like Amazon, Microsoft, and others.

  33. @Tim the Coder

    Yawn. You’ve just invented the mini-cab.

    Self-driving cars are only cost-effective when it costs too much to have a human drive them. The point is to save money by having 10 maintenance guys instead of 1000 drivers.

    Shortly, people will talk about self-driving cars having to have dedicated routes

    This is how the large company will force people to use its product instead of buying their own cars. You evil gas-guzzling planet-wreckers get one lane, the virtuously-green-self-driving cars get four.

  34. I heard Shashi Vermi CTO at Transport for London speak on this subject recently. He is highly skeptical about the subject of autonomous vehicles on public roads, his view:
    – Humans accept humans kill people, but machines killing humans is another matter and puts autonomous vehicles on a different safety threshold that’s incredibly hard to meet.
    – How will the computer decide whether or not to stop when somebody steps out in front of it, if it always stops the car will never go anywhere in busy locations, and if it doesn’t stop people are going to start asking difficult questions about whether it is okay for a computer programme to run people over?
    – Autonomous vehicles are already out there in controlled environments e.g. they are used in mining operations in Australia, but these are environments where human interaction with the vehicle is highly restricted. You cannot restrict people from roads of London like you can in a remote mine in Australia.
    – The costs of resolving the outstanding issues are prohibitive; manufacturers simply aren’t making process on the most intractable issues. As mentioned the technology to have autonomous vehicles has already been delivered it’s meeting the threshold for public roads that’s going to take shitloads of effort and money.

  35. By the sound of it, there’s a great deal of thinking inside the box, here. The idea of an autonomous car is ludicrous. Makes about as much sense as putting autonomous trains on the rail network or flying autonomous airliners in & out of airports. And it would be entirely possible, right now, to have driverless trains & pilotless aircraft.
    So try thinking outside the box. What you might want a driverless car to do & how that might be accomplished. Instead of trying to replicate a human car driver.

  36. I have no particular insight into the technology involved

    That would be the problem with most of the posts on this topic anywhere.

    I’m not going to waste time reiterating variations on “no, that’s not true, and yes, they’ve already considered that”. I’ll just repeat my offer: 5:1 that self-driving cars that are good enough (defined as having an accident severity rate per miles driven equal to or less than human-driven cars) are available for commercial sale within seven years.

  37. Simple example. Positional accuracy:
    Sure. A few meters if you’re using the crap GPS in your mobile phone. Get yourself a dedicated GPS receiver & some decent software & computing power & acccuracy of a foot isn’t hard.
    But that’s still using the GPS constellations out in geosync. Why? Turn the GPS off & you’re phone still knows where it from the relay masts. It wouldn’t be at all hard to use 4G transmissions from several known locations to triangulate positional accuracy down to inches.Or just sow the country with dedicated position transmitters. Stick them on the top of streetlamps, if you like. Now you can be down to millimeters. Better than any human eyesight could estimate.

  38. @Daniel – “I’ll just repeat my offer: 5:1 that self-driving cars that are good enough (defined as having an accident severity rate per miles driven equal to or less than human-driven cars) are available for commercial sale within seven years.”

    Me too. I also would not waste my time arguing with an uneducated early model isolated Eskimo. Particularly an elderly one that only knows how to post when his dial up t’internet is in sync with the nearby hungry wolves that are baying at the moon.

  39. “… with the parts carrying out the safety-critical functions kept well separate from the bits that run the entertainment system…”

    Well that’s a relief. But according to what we read in the popular press, current manufacturers of (non-autonomous) vehicles haven’t worked out that that is an important issue.

    Doesn’t raise my hopes much, I have to admit.

  40. ” It wouldn’t be at all hard to use 4G transmissions from several known locations to triangulate positional accuracy down to inches…”

    Well it would be pretty hard where I live, as there is no 4G signal available – as there isn’t in quite large areas of the the country. 90% of the population covered does NOT equal 90% of the land area, and certainly not 90% of the roads.

  41. @Jonathan: Self-driving cars are only cost-effective when it costs too much to have a human drive them. The point is to save money by having 10 maintenance guys instead of 1000 drivers.

    True that a mature AI might (maybe!) be cheaper than a human.
    But the labour cost of an Uber is only a small part of the total cost, including amortised capital.
    And the human driver helps reduce the cleaning bill between hires…

    My point is that the business model has existed for >50 years, so self-driving cars, if ever possible, doesnt change that. Self-driving minicabs are still minicanbs, though maybe with better conversation.

    Self driving personal auto is still a personal auto. Good for getting home from t’pub.

    But I doubt a true self driving car will ever appear, there are huge technical diifficulties, software being noted for freedom from errors, no?
    Also, as was pointed out, the manufacture carries all liability. Guess who pays?

    Regarding 4G triangulation: no mobile operator would accept any such life-critical usage. If used without such support, it would then be ‘upset’ by the mobile operators in their own self-defence.

  42. For motorway driving (level 4), we’re almost there. The problem will be massively increased demand. Thousands more people will accept jobs half-way around the M25, knowing they can sleep most of the way. Holidaymakers will sleep-drive overnight to the south of France, bringing as much luggage as they want, and not having to drive to the airport’s long-term car park at 4am or worry about car hire rip-offs.

    As your car drives onto the motorway on-ramp, a little green light bleeps and the car’s computerised voice says “now entering autonomous driving zone”. This should prove very profitable for toll road operators like SANEF.

  43. Tim replies to one of my questions (In-vehicle: how powered and cooled?): “Power generation in a car with a petrol or diesel engine isn’t a problem, nor is cooling. With an electric car it’s more tricky, but if the power problem is solved for propulsion it won’t be a problem for running a computer either.”

    A typical-ish nice car has an engine that provides peak power of around 100kW (133bhp).

    My very broad assumption for an AI computer system good enough to drive a car at Level 5 (any surfaced road anywhere with no external help beyond the current type of Satnav) is some several PetaFLOPS of computational power supported currently by electricity consumption of approaching 1MW.

    Now, that 1MW is the equivalent of 10 of those typical-ish nice cars running at full power, all the time. So around 30 such cars powered to typical cruising speeds, or a similar consumption ratio for lowish and very inefficient urban speeds.

    Is Tim content to argue for a 30ish-times increase in fuel costs – over cars driven by human intelligence. If not, what does he believe is likely to be the power requirements for computational (on-board or off-board) Level 5 provision?

    Whilst I am offering no evidence (we have none) for the power requirements I claim are likely to be necessary for Level 5 autonomous vehicles, neither is Tim. I (at least) am offering a power requirement below that for which we actually do have computers. And below that for computers (pretty much the best we have) on which no one has yet demonstrated anything approaching Level 5 automotive vehicle performance, nor offered any sort of demonstration of other Artificial General Intelligence (AGI).

    If the algorithms are ever going to exist for Level 5 practical (ie cost effective) replacement of human drivers, surely (and long before that) we are going to have the algorithmic performance without the cost-effectiveness. This most likely would be using off-board computation and by routing the input data feeds (visual, radar (?really?), lidar, the rest) to those ‘cloud’ computers. With such demonstration, we would be able to see at least some sort of defined (if not absolutely clear) path to such a widely available capability.

    Best regards

  44. My very broad assumption for an AI computer system good enough to drive a car at Level 5 (any surfaced road anywhere with no external help beyond the current type of Satnav) is some several PetaFLOPS of computational power supported currently by electricity consumption of approaching 1MW.

    In which case I stand corrected!

  45. “But the labour cost of an Uber is only a small part of the total cost, including amortised capital.”

    Are you absolutely certain about that? It’s easy enough to confirm; Uber recently released ride/payment data (which proved the gender pay “gap” is bollocks, by the way). I think you’ll be surprised.

    And the human driver helps reduce the cleaning bill between hires…”

    How? If you throw up in the car, he/she adds a surcharge to the fare. What’s that, 20 lines of code to perform the IF/THEN test?

    Regarding 4G triangulation: no mobile operator would accept any such life-critical usage. If used without such support, it would then be ‘upset’ by the mobile operators in their own self-defence.

    Remind me; when you dial 911/999/000, which organisations are involved in providing that mission critical service?

  46. I’d be more of a cynic about this if the tech didn’t already exist pretty much in action. My understanding is that the the Tesla autopilot is so good that the main problem is keeping the driver sufficiently interested they don’t start reading the paper. My memory of the “trans bankrobber crashes robot car into drugged up jaywalker” episode is that when the accident video was released, it became apparent that had the car been jointly driven by Sterling Moss and Mother Teresa they would still have driven over said jaywalker. I presume there have been remarkably few other incidents, given the media attention that one got.

    All that said, I’d be very loathe to share any road space with an autonomous Renault. A friend of mine bought a newish VW with collision avoidance, after two full on emergency stops for no apparent reason in the first two weeks of ownership he decided that the risk of having a 44 ton truck drive over the top of him while braking to avoid a butterfly was too great – the dealership agreed that the car was “faulty” and allowed a full refund, so clearly some manufacturers have a long way to go (and I think collision avoidance may be due to become mandatory on new cars in the not too distant future!).

  47. @Jonathan: Self-driving cars are only cost-effective when it costs too much to have a human drive them. The point is to save money by having 10 maintenance guys instead of 1000 drivers.

    I am reminded of the problem of machine translation. Quite good-enough for specialist uses results were possible by the early 60s, but before you could translate you needed the text in machine readable form. Back then it was not economic, human translators were cheaper to employ than an accurate copy typist.

  48. theProle

    “I’d be more of a cynic about this if the tech didn’t already exist pretty much in action. My understanding is that the the Tesla autopilot is so good that the main problem is keeping the driver sufficiently interested they don’t start reading the paper.”

    Isn’t this simply a reflection of the challenge to get from a 99% system to a 99.999% system? It’s easy to automate easy things, it’s many orders of magnitude harder to get it to work on complex things. So, introducing automation that works in ideal conditions has already happened, but all this achieves is people turning off and not being positioned to cope when the complex situation occurs?

    What lessons does aviation have to teach?

Comments are closed.