12 ms·

It feels there's a mountain of hype around autonomous vehicles. I think the core challenges to a fully autonomous vehicle (with no human backup), are still far

It feels there's a mountain of hype around autonomous vehicles. I think the core challenges to a fully autonomous vehicle (with no human backup), are still far from solved. They need to reliably deal with an almost infinite number of edge and corner cases, each quite different from the last. For example: - communication with other human drivers. In London, this is required all the time, like when parked cars block the road, allowing just one car through. Or traffic light out of action, so you negotiate with other cars using hand gestures, light flashes etc - endless roadworks, that change what lane you're allowed to go on, turn a two way road to one way road. - random debris on road. Plastic bag - safe to drive through, wooden plank - safe, plank with nail - not safe. - loss of GPS, mobile data, or both (again, surprisingly frequent) - making way for emergency vehicles (sometimes need to drive into lane you're not normally allowed to go, I.e. Bus lane, pavement) - policeman coordinating traffic So far, I haven't found any evidence of autonomous cars dealing with the above. If anyone has, please post.

larrik 10y ago
I'm not sure what your point is. Fully autonomous is still in progress, and still has work to do.
- koheripbal 10y ago
  The point is that we are far far further away than commonly accepted. The day a computer can distinguish between a plastic bag in the road and a child is the day I'd support them on the road.
  - woofyman 10y ago
    Or if you see a ball cross the road there are strong odds a child might be chasing after it.
  - zoul 10y ago
    Do computer drivers have to be perfect, or is it enough to be much better than human drivers?
    - ZeroFries 10y ago
      Perfection is of course, impossible, the best we can hope for is progress asymptotically close to ideal. The minimum bar should be performance surpassing human performance, but the details do matter. For example, a 99.99% safety rated car which almost never crashes but randomly explodes once every 10,000,000 uses would probably not be accepted.
  - mypalmike 10y ago
    The day a human can distinguish between a plastic bag in the road and a child is the day I'd support them on the road.
    - mypalmike 10y ago
      My mother once almost killed a child who ran into the road. The kid was on the side of the road and suddenly ran into the street. She did not see him, and he ran into some sharp piece of metal on the car (this was in the early 60s). He bled profusely, and the neighborhood people came out and almost became violent with her, blaming her for the incident. If she had been in an automated car, it would have almost certainly have stopped before the collision. And what's wrong with slowing/stopping for a plastic bag? How often do plastic bags really sit in the street shaped like babies?
- achamayou 10y ago
  His point is that the remaining 5% of a problem can take considerably longer to solve than might initially be anticipated by looking at the quick progress made on the first 95% of it. A textbook example of that is voice recognition. It's also not particularly clear that the costs will drop rapidly to zero or even to a level that leads to mass adoption (see air travel at Mach 1+, Segways...).
  - baq 10y ago
    Your examples aren't very good, sorry. We're talking about adding AI, autopilot, autonomous capabilities, whatever you want to call it to already existing hardware. It's mostly a matter of software and you're a living example that it can be done, in 20W, with nothing more than stereo cameras and mics. If I had to compare, I expect the change to be more akin to what Excel has done to paper spreadsheets.
    - thecrazyone 10y ago
      You're oversimplifying it and being a little pessimistic. Why? Because there's an order of magnitude or more difference in the complexity of AI of cars and AI that a human uses to navigate
    - achamayou 10y ago
      Current voice processing tech is still very much worse than an average human, no matter how fancy the microphones and how much power it's allowed to consume from the grid. Good sensors aren't cheap, the optical and mechanical elements aren't likely to drop in price soon. Only the computing power is getting cheaper, but harnessing it isn't particularly easy (we're not getting better at programming as quickly as we're making faster chips). If you need another example, consider automated cleaning robots. It's a considerably simpler problem to solve, and they're still only a marginal product. At the moment, autonomous tech is only really used in well-controlled industrial environments. It will surely break out of there progressively, but chances are it will take a while.
      - baq 10y ago
- walrus1066 10y ago
  That people are talking as if it's just round the corner, that taxi drivers and truckers are about to become obsolete. Whereas I understand the technical challenges are much greater than is stated, it may even be they won't be all overcome.
  - ajross 10y ago
    But so what? Full autonomy needn't even be a business goal of Waymo. They just want full deployment. Having a system in place which reduces or eliminates many/most of the accident root causes would be worthwhile to install on any vehicle on the road even if it technically had a human driver. The "goal", such as it is, is to have a technology worth using. It's not to transcribe whichever comic book utopia someone has in mind.
    - obmelvin 10y ago
      Their stated goal is indeed full autonomy.
  - munificent 10y ago
    It's not that taxi drivers and truckers are going to become obsolete tomorrow. It's that that happening is such a huge cultural and economic change that we need to start planning for it today if we're going to not destroy the livelihood of millions of people when those industries do become obsolete in five or ten years.
    - Fricken 10y ago
      About 20 million jobs disappeared in America last year, although there was a net gain in jobs overall. This is going on all the time. So if all of America's 11 million professional driving jobs were to disappear over the span of a decade, it would be a blip, not an apocalypse.
  - agildehaus 10y ago
    Tesla is launching Autopilot 2.0 (in stages) starting in January. It's still NHTSA level 3, not 4 or 5, but it's a major step up from pretty much anything currently available and will be in consumer's hands. Even more so when the Model 3 comes. I think you'll be surprised how fast this technology will exit the labs. Just that Google has decided to brand around it is a big indicator that they're growing confident.
CardenB 10y ago
A lot of these problems can be solved with a combination of telemetry, limited scope rollout, and an emergency abort procedure.
thebspatrol 10y ago
In terms of communication, I think the end game is a Las Vegas/Audi style centralized system of traffic negotiation and communication. But that's an obvious chicken or egg problem that has unrealistic solutions like making sure everyone starts driving a networked car at all once.
walrus1066 10y ago
Regarding emergency vehicles, found this, but haven't seen videos of it actually in action http://www.freepatentsonline.com/20160252905.pdf http://www.freepatentsonline.com/20160252905.pdf
notatoad 10y ago
Do you really think the companies building autonomous cars haven't thought about all these things? They may not have solved them yet (If they had, we've probably see their products on the road) but I'm sure it's in progress.
- monk_e_boy 10y ago
  I imagine in an emergency the car would show a set of options on a screen and you click one and the car will do it. For example getting out of the way of emergency vehicles on a blocked road. It will ask you what to do and if you tell it to drive up the pavement or into a bus lane or cross to the wrong side of the road -- the car will just do it. I do wonder about all the non-verbal human-to-human communication. But a friendly baxter type screen with eyes that move (and messages that scroll across) would solve that. We have to remember that these cars are all feeding unusual situations back to the google BRAIN that is learning. It'll watch us react to an emergency and learn. It'll see that 20 or 30 times a day and generalise a response.
  - walrus1066 10y ago
    I think it's fine if there's a human in the car, simple, they just take control to manage the incident. But to truly disrupt our lives (24 hour robot trucks, autonomous taxi fleets ), which is what all these companies are aiming for, the car needs to be able to drive without a human.
    - Retric 10y ago
      Having a human on call to remote into a Car is still a viable option in most of the US. This is not a good fit for sub second response times or tunnels, but it vastly reduces the endless .0001% edge cases. ED: Not hitting kids is manditory, not blocking traffic is mandatory, parking in every random stop is optional as long as their are safe places it can drive to.
      - walrus1066 10y ago
londons_explore 10y ago
These issues could mostly be solved by the car having a "remote control" where an expert in a remote immersive environment could take manual control of the car for a few seconds to resolve whatever the problem is. Now you've reduced all your problems down to simply "identify problem and alert a remote human". If each of those problems only happen once per day, and they are resolved in 30 seconds, it's going to cost only about $30 per car per year to have that remote expert on hand.
- walrus1066 10y ago
  That's an option, but London has loads of signal black holes, or the speed is just not good enough. Infinitely worse in the country, if we want autonomous lorries. I think we'd need a huge infrastructure upgrade to make life easier for autonomous cars. It'd have to be >100 Mb/s mobile data, with extremely good coverage (for the recovery mechanism you mention). Think GPS would need to be replaced too, the signal is way too weak and unreliable. We'd have to embed digital markers/emitters alongside lane markings, traffic lights, signs etc, so they're easier to deal with.
  - londons_explore 10y ago
    No need for high speed - all these issues can be resolved by inching forward or back to get the car back onto the open road. (although obviously high speed is a bonus) Connectivity you see is from a single mobile phone. A car could have a better antenna, service from all the mobile networks rather than just one, wifi fallback and a backup satellite uplink. If all those fail, you dispatch an engineer to find the car wherever it's stuck and add a solar powered LTE repeater. If you have a map of the whole service area and know where you started and have a camera looking out the window the whole time, there is no need for GPS really.
    - EdHominem 10y ago
      It might even be the sort of thing where you tap the video of the street ahead to tag the bag as passable and the car then slowly drives through it. No remote guidance per-se. Considering these are car companies doing this, true remote control is not a good idea. The industry response to the car hacks was pathetic. Only Tesla reacted properly. (And probably Google...)
iraklism 10y ago
I get that we have to be open with tech. After all most tech seemed "impossible" "unrealistic" and "unfeasible" even a couple of years before they go mainstream. But you have touched some use cases that are more frequent than most people think. London is a great example. My hometown Athens is much more closer to London as far as driving is concerned as opposed to LA or NY. I guess many other metropolis also. How will AI brand A communicate with AI brand B about life critical decisions ? Hey we can either both crash to each other or kill that kid on the pavement. ( just an example , I don't want to go to the morality of the situation) I assume there is going to be a "standard " . But have a look at your everyday tech. Standards are not easy. There are still WiFi routers/repeaters that cannot talk to each other nicely, just to give an example. Usually in life critical systems you don't have interoperability issues because there is no interoperability. Huge fail-safe systems from a single vendor.
- walrus1066 10y ago
  I can see communication with other autonomous cars happening, but with humans is another question altogether. We enter the realms of 'general ai' here, which is way out of scope of our ML tools.
  - ThomPete 10y ago
    Is that really true though? You don't use your general AI to see someone extending their hand in traffic ex. General AI is the ability to both drive car, score a girl, cook food, clean your room etc. I think you would be surprised how little general AI have to do with this.
    - walrus1066 10y ago
      Sure, not Turing test level AI, but the car needs not to just correctly interpret what the other human driver is communicating, it too needs to communicate back, sometimes assertively. Sometimes you need to actually talk to the other driver. Anecdote: I was driving down a long, narrow country road, lots of turns, space only for 1 car, land rover coming in opposite direction, and another car behind me. We had to verbally negotiate who reverses back 50m to let the others go past. I cannot think how an ML algo could resolve this situation, even if it could, how would it communicate it's intentions to the humans, when flashing your lights isn't enough?
woofyman 10y ago
And that's just a partial list: deep pot holes, construction zones, standing water an inch deep or is it a foot deep?
- lifeformed 10y ago
  How do humans tell how deep water is?
  - trome 10y ago
    Humans generally look at it, make an informed guess and drive accordingly. The recent Tesla accident is a great example of how even the best optics can be blown out by environmental factors like sun, whereas a human would likely readjust their vantage point, lower the sunshade, put on sunglasses or even pull over if they are unable to cope. Comparatively, the car did not attempt to do any of this and actually didn't even realize its cameras were in a state of failure.
    - gsam 10y ago
      Which says a lot about relying specifically on optics. Readjusting vantage points from an autonomous vehicle is as simple as having multiple sensors and deciding to rely more on a particular sensor in particular conditions. In particular, Google has some of the most expensive arrays of sensors.
brosky117 10y ago
The page about technology says, "Our sensors observe that the cyclist has extended their left arm. Detecting the cyclist's hand signal, our software predicts that the cyclist will move to the left side of the lane." This suggests to me that they're trying to deal with situations like the ones you've brought up.
- moxious 10y ago
  There is a difference between "trying to deal with" and "product on sale next month". These cars are overhyped. They've been just around the corner for the last 10 years. They are the Duke Nukem Forever of Silicon Valley.
  - criley2 10y ago
    >These cars are overhyped Not really, they're correctly hyped. The concept is revolutionary. > They've been just around the corner for the last 10 year B-S. Holy revisionism BS!! We were NOT talking about self-driving cars in 2006! There was no company seriously designing fully autonomous cars in 2006! It was fully in the realm of science fiction, no investment, no major players, nothing. And you make it sound like it was supposed to be released 10 years ago! Vaporware! They told us we'd have AI cars in 2006! It never came! <---- False reality that never occurred. Why do you feel the need to lie about recent history? Shouldn't you observe that if you have to make a nasty lie to support your argument, maybe your argument is wrong? You're definitely parroting the contrarian point here, but you're being very low-effort about it. Contrarian can be cool but please try harder.
    - ZeroFries 10y ago
      https://en.wikipedia.org/wiki/History_of_autonomous_cars https://en.wikipedia.org/wiki/History_of_autonomous_cars
      - ssalazar 10y ago
        A link to a 15 page wikipedia article is not a rebuttal.
swamp40 10y ago
They don't need to make them foolproof. They just have to make them safer than human drivers. And these days with everyone swerving as they text, the bar isn't exactly set very high.
- legolas2412 10y ago
  They need to make it safer than average driver being adviced by driver assistance systems. That system can fix the negatives of human drivers without affecting the advantages of humans over machines
BurningFrog 10y ago
A lot of that will be best solved by making a lot of other actors communicate directly with the car. Traffic lights will send out the same messages they send to humans by red/green lights to the cars using radio. Emergency vehicles, roadworks, traffic signs etc will do the same. Or so I assume, because it makes so much of sense. I know there is vehicle-to-vehicle communication in the works: https://en.wikipedia.org/wiki/Vehicle-to-vehicle https://en.wikipedia.org/wiki/Vehicle-to-vehicle
- walrus1066 10y ago
  Yes, I think you're right. We'd need to update a lot of our infrastructure to make it friendlier for these cars. I should also add: - very fast mobile data, with no blackspots, so a remote driver can take over a car that's confused by a situation - a replacement for GPS, it's not reliable enough, the signal is far too weak - digital signal beacons embedded in lane markings, and edges of the road, and on traffic cones for roadworks
  - trome 10y ago
    Why should local municipalities foot the bill for adding beacons to stop lights, etc? I get trying to make infrastructure friendly towards self driving cars, but the general public should not be footing the bill for someone's life choice to buy such a vehicle.
    - walrus1066 10y ago
      Because if we do resolve these issues, our roads will be much, much safer with autonomous cars. 2k people die on the roads each year in the UK, multiples more sustain life long injuries. I was in an accident, which could have been prevented were it not for the 1 second it takes for a human to react to a situation.
    - BurningFrog 10y ago
      This is important traffic infrastructure, just like visual traffic lights and asphalt. That said, it can be paid for my many actors.
Angostura 10y ago
A lot of these problems can be solved by simply making them fly. ... that's right, now you know what the driving force behind all those flying cars in sci-fi was. It's simply a way to separate the autonomous and old-style vehicles.
- elmar 10y ago
  you are right, driverless cars are so passe the future are self-flying vehicles.
- jugchock 10y ago
  This seems to have been said in jest but the main thing keeping flying cars in sci-fi and out of the real world is the human element. Most people have a hard enough time driving in two dimensions. Adding a third would be chaos without the aid of our robot overlords.
  - dragonwriter 10y ago
    > This seems to have been said in jest but the main thing keeping flying cars in sci-fi and out of the real world is the human element. I thought it was that winged flight makes for crappy cars and wingless, vectored thrust flight is horribly fuel hungry. (Sci-fi versions are usually wingless and either have magic drive systems or just ignore the fuel issue.)
bhntr3 10y ago
My scuba instructor shared this piece of wisdom: "You don't have to swim faster than the shark. You just need to swim faster than your buddy." People get this stuff wrong all the time. Especially when they're drunk, tired or texting. These cars won't be perfect but I think pretty good and conservative under uncertainty will be better than human, drunk and overconfident.
- ticviking 10y ago
  That's a key point here. The cars don't have to be perfectly safe, just safer than the average driver.
  - woofyman 10y ago
    It has be better than the best human driver or they'll get sued out of existance. You can't sell a product that you know will hurt or kill a certain number of people.
    - deadbunny 10y ago
      I'm pretty sure the car industry has been doing that for years.
      - woofyman 10y ago
        They get sued if they knowingly sell a defective product.
    - ageofwant 10y ago
      Sorry no. Cars kill thousands of people every year, and will continue to do so. Car companies are doing just fine. Power companies literally kill millions every year, a-ok. McDonalds, Coca Cola ? ditto.
      - woofyman 10y ago
        Drivers kill, not the cars. That's totally different from an autonomous car killing someone. Humans make mistakes. That's different from a defective product.
LinuxBender 10y ago
I agree with all of this. I would also add; that I will stay consider these vehicles to be unsafe until the manufacturers can prove to be beyond a shadow of doubt that I or the local computer are the ones piloting the vehicle. When a random person can remotely take control of a vehicle; whether it be the windshield wipers, or full control of the vehicle, then it is not safe to contain a power source in the vehicle near a public road. I don't want 4chan driving me into another vehicle.
- andrewljohnson 10y ago
  I'd consider autonomous cars to be more safe than a car I drive when statistically I'm less likely to get hurt or die. If there was some chance my car could get hacked, or other weird vulnerabilities that robocars had that traditional cars did not, I would still want to be in a robocar if my chance of dying was cut in half.
  - lazaroclapp 10y ago
    Sure. But you are assuming a good way to estimate that chance. Say you base it on previous performance: in the last 5 years, twice as many people have died driving non-self-driving cars compared to those riding self-driving cars. Then, the very next day, a massive hack / cyberattack happens that causes 10 million fatal accidents...
  - brazzledazzle 10y ago
    This is going to be one of those things where a lot of people have trouble quantifying the actual risk because the fear when things are out of your control is so powerful. Like flying.
    - LinuxBender 10y ago
      Not to mention, the live hacks of moving vehicles have already occurred multiple times. Thankfully, there are not yet enough skiddie tools and the bar to entry is just high enough that mass casualties have not yet occurred. Give it a few more months or maybe even another year before your car is just another IoT that can be controlled remotely. I do not believe it would be irrational to consider the risk of 100k cars driving into oncoming traffic at the same time. That is a lot more damage and chaos than a few little 747's could do. Am I incorrect? My only questions are 1) will the skiddie tools be on github first, or will someone have to buy/leak them and repost on github? 2) If a minor executes the scripts and lives are taken; and assuming you can determine who ran them, will they be tried as adults and will the verdict or punitive actions make up for the damages?
yohui 10y ago
https://www.wired.com/2016/12/google-self-driving-car-waymo https://www.wired.com/2016/12/google-self-driving-car-waymo > By early 2015, Fairfield thought they were getting close: The cars were clocking full days at Google’s testing grounds without needing human takeovers. They had driven 1.2 million miles on public roads and could pull over for emergency vehicles. They could detect and brake for squirrels, and read hand signals from cops and construction workers. They knew when to honk politely (Just making sure you see me here) and when to blare it (You’re about to slam into me!). Then the team spent the next year putting another 800,000 miles on the fleet’s collective odometer, to fine tune everything. And it gave 10,000 rides to employees and “guests.”
- walrus1066 10y ago
  The car-human interaction bit isn't from a quote, is there any evidence of this actually happening in the field? Would be great if there was a video or something.
  - marwatk 10y ago
    You can see one example at the end of the video on their technology page where the cyclist can't make up his mind. It recognizes the left turn sign from the rider (you can see the sybmol change when he extends his arm). https://www.youtube.com/watch?v=fbWeKhAPMig https://www.youtube.com/watch?v=fbWeKhAPMig
    - StRoy 10y ago
      This kind of predictable reactions to specific human gestures.. I fear it will have great uses for certain type of crimes. What will you do if someone stops you with the purpose of assaulting _for eg, a robbery_ you and the car refuses to run the dude over because he made a gesture and is now in front of your car while his buddies are waiting aside? There are good reasons why things like these happened: > The law of some states, such as Louisiana, explicitly lists a killing in the course of defending oneself against forcible entry of an occupied motor vehicle as a justifiable homicide. While I'm interested in self driving cars that still feature manual controls you can override, like Tesla, I have no interest in a future controlled by Google's AI without my input.
agumonkey 10y ago
All these are old "problems", either solved or in progress. Something I've never seen adressed is how SDV fleets behave when change occurs. Are there chaotic oscillation patterns that emerge from AI trying similar avoiding behavior. There's research in massive multi agent behavior, but I've never seen it tested on the field with SDV makers actual models.
spuz 10y ago
I think a good way to think about the last 5% automation problem is to look at other examples where automation has replaced human work. In every case that I can think of, the work of the human was never 100% replaced. Automated check outs at the supermarket still require a number of attendants. An automated assembly line still needs factory workers. Automated rail systems are still controlled by humans in a control room. You can also look at how the environment has been adapted to better suit the new technology. For example to allow big heavy horseless carriages roll along roads we replaced dirt with tarmac (asphalt). Driverless cars will never (until we have general AI) be able to fully replace a human driver. However we don't need to get to 100% for the technology to become useful. It might be that driverless cars will be restricted to only driving on certain mapped roads. Or every car's movements are tracked in a central control room and if a problem is detected either by the car or the passenger it could be controlled remotely. Or let's say a GPS sensor breaks an actual human driver might be issued to come and rescue you. Lastly we might even accept a certain failure rate. Society is willing to tolerate failure of technology to a certain degree in cars and planes today. All we need is is for the technology to make us all better off on average for it to be successful.
teknoxjon 10y ago
I'd imagine all the GPS data could be loaded into the system beforehand and updated when there is a consistent signal. The same way cars now deal with GPS.
nialv7 10y ago
> They need to reliably deal with an almost infinite number of edge and corner cases, each quite different from the last. It doesn't need to be perfect. It just needs to be better than average human drivers. Sure there will be cases it can't handle. But the same is much more true for human.
- dilemma 10y ago
  Red herring. It doesn't need to be perfect; it needs to handle an infinite number of edge cases. Humans can, machines cannot.
  - nialv7 10y ago
    Strictly speaking, it needs to handle an infinite number of edge cases _in a very limited amount of time_ Machines might not be able to, humans definitely can't.
sandworm101 10y ago
>> communication with other human drivers. I would shorten that to "other humans", to specifically include pedestrians. That 1/2-second eye contact between driver and human is key. Would anyone here dare walk in front of a car without that little nod from the driver? The other edge case, one that is never discussed, is how these things expect to deal with off-road driving. I regularly use car ferries. Getting on and off those, and negotiating the terminals, stumps many human drives.
Shivetya 10y ago
I am still waiting for the at night in full on rain demos on unmarked roads or such.
redwood 10y ago
What you're missing is "how do humans cope with these scenarios?" -- by either cautiously slowing down or by following those humans ahead or both. Machines can do the same

larrik 10y ago

I'm not sure what your point is. Fully autonomous is still in progress, and still has work to do.

koheripbal 10y ago
The point is that we are far far further away than commonly accepted. The day a computer can distinguish between a plastic bag in the road and a child is the day I'd support them on the road.
- woofyman 10y ago
  Or if you see a ball cross the road there are strong odds a child might be chasing after it.
- zoul 10y ago
  Do computer drivers have to be perfect, or is it enough to be much better than human drivers?
  - ZeroFries 10y ago
    Perfection is of course, impossible, the best we can hope for is progress asymptotically close to ideal. The minimum bar should be performance surpassing human performance, but the details do matter. For example, a 99.99% safety rated car which almost never crashes but randomly explodes once every 10,000,000 uses would probably not be accepted.
- mypalmike 10y ago
  The day a human can distinguish between a plastic bag in the road and a child is the day I'd support them on the road.
  - mypalmike 10y ago
    My mother once almost killed a child who ran into the road. The kid was on the side of the road and suddenly ran into the street. She did not see him, and he ran into some sharp piece of metal on the car (this was in the early 60s). He bled profusely, and the neighborhood people came out and almost became violent with her, blaming her for the incident. If she had been in an automated car, it would have almost certainly have stopped before the collision. And what's wrong with slowing/stopping for a plastic bag? How often do plastic bags really sit in the street shaped like babies?
achamayou 10y ago
His point is that the remaining 5% of a problem can take considerably longer to solve than might initially be anticipated by looking at the quick progress made on the first 95% of it. A textbook example of that is voice recognition. It's also not particularly clear that the costs will drop rapidly to zero or even to a level that leads to mass adoption (see air travel at Mach 1+, Segways...).
- baq 10y ago
  Your examples aren't very good, sorry. We're talking about adding AI, autopilot, autonomous capabilities, whatever you want to call it to already existing hardware. It's mostly a matter of software and you're a living example that it can be done, in 20W, with nothing more than stereo cameras and mics. If I had to compare, I expect the change to be more akin to what Excel has done to paper spreadsheets.
  - thecrazyone 10y ago
    You're oversimplifying it and being a little pessimistic. Why? Because there's an order of magnitude or more difference in the complexity of AI of cars and AI that a human uses to navigate
  - achamayou 10y ago
    Current voice processing tech is still very much worse than an average human, no matter how fancy the microphones and how much power it's allowed to consume from the grid. Good sensors aren't cheap, the optical and mechanical elements aren't likely to drop in price soon. Only the computing power is getting cheaper, but harnessing it isn't particularly easy (we're not getting better at programming as quickly as we're making faster chips). If you need another example, consider automated cleaning robots. It's a considerably simpler problem to solve, and they're still only a marginal product. At the moment, autonomous tech is only really used in well-controlled industrial environments. It will surely break out of there progressively, but chances are it will take a while.
    - baq 10y ago
walrus1066 10y ago
That people are talking as if it's just round the corner, that taxi drivers and truckers are about to become obsolete. Whereas I understand the technical challenges are much greater than is stated, it may even be they won't be all overcome.
- ajross 10y ago
  But so what? Full autonomy needn't even be a business goal of Waymo. They just want full deployment. Having a system in place which reduces or eliminates many/most of the accident root causes would be worthwhile to install on any vehicle on the road even if it technically had a human driver. The "goal", such as it is, is to have a technology worth using. It's not to transcribe whichever comic book utopia someone has in mind.
  - obmelvin 10y ago
    Their stated goal is indeed full autonomy.
- munificent 10y ago
  It's not that taxi drivers and truckers are going to become obsolete tomorrow. It's that that happening is such a huge cultural and economic change that we need to start planning for it today if we're going to not destroy the livelihood of millions of people when those industries do become obsolete in five or ten years.
  - Fricken 10y ago
    About 20 million jobs disappeared in America last year, although there was a net gain in jobs overall. This is going on all the time. So if all of America's 11 million professional driving jobs were to disappear over the span of a decade, it would be a blip, not an apocalypse.
- agildehaus 10y ago
  Tesla is launching Autopilot 2.0 (in stages) starting in January. It's still NHTSA level 3, not 4 or 5, but it's a major step up from pretty much anything currently available and will be in consumer's hands. Even more so when the Model 3 comes. I think you'll be surprised how fast this technology will exit the labs. Just that Google has decided to brand around it is a big indicator that they're growing confident.

CardenB 10y ago

A lot of these problems can be solved with a combination of telemetry, limited scope rollout, and an emergency abort procedure.

thebspatrol 10y ago

In terms of communication, I think the end game is a Las Vegas/Audi style centralized system of traffic negotiation and communication. But that's an obvious chicken or egg problem that has unrealistic solutions like making sure everyone starts driving a networked car at all once.

walrus1066 10y ago

Regarding emergency vehicles, found this, but haven't seen videos of it actually in action http://www.freepatentsonline.com/20160252905.pdf http://www.freepatentsonline.com/20160252905.pdf

notatoad 10y ago

Do you really think the companies building autonomous cars haven't thought about all these things? They may not have solved them yet (If they had, we've probably see their products on the road) but I'm sure it's in progress.

monk_e_boy 10y ago
I imagine in an emergency the car would show a set of options on a screen and you click one and the car will do it. For example getting out of the way of emergency vehicles on a blocked road. It will ask you what to do and if you tell it to drive up the pavement or into a bus lane or cross to the wrong side of the road -- the car will just do it. I do wonder about all the non-verbal human-to-human communication. But a friendly baxter type screen with eyes that move (and messages that scroll across) would solve that. We have to remember that these cars are all feeding unusual situations back to the google BRAIN that is learning. It'll watch us react to an emergency and learn. It'll see that 20 or 30 times a day and generalise a response.
- walrus1066 10y ago
  I think it's fine if there's a human in the car, simple, they just take control to manage the incident. But to truly disrupt our lives (24 hour robot trucks, autonomous taxi fleets ), which is what all these companies are aiming for, the car needs to be able to drive without a human.
  - Retric 10y ago
    Having a human on call to remote into a Car is still a viable option in most of the US. This is not a good fit for sub second response times or tunnels, but it vastly reduces the endless .0001% edge cases. ED: Not hitting kids is manditory, not blocking traffic is mandatory, parking in every random stop is optional as long as their are safe places it can drive to.
    - walrus1066 10y ago

londons_explore 10y ago

These issues could mostly be solved by the car having a "remote control" where an expert in a remote immersive environment could take manual control of the car for a few seconds to resolve whatever the problem is. Now you've reduced all your problems down to simply "identify problem and alert a remote human". If each of those problems only happen once per day, and they are resolved in 30 seconds, it's going to cost only about $30 per car per year to have that remote expert on hand.

walrus1066 10y ago
That's an option, but London has loads of signal black holes, or the speed is just not good enough. Infinitely worse in the country, if we want autonomous lorries. I think we'd need a huge infrastructure upgrade to make life easier for autonomous cars. It'd have to be >100 Mb/s mobile data, with extremely good coverage (for the recovery mechanism you mention). Think GPS would need to be replaced too, the signal is way too weak and unreliable. We'd have to embed digital markers/emitters alongside lane markings, traffic lights, signs etc, so they're easier to deal with.
- londons_explore 10y ago
  No need for high speed - all these issues can be resolved by inching forward or back to get the car back onto the open road. (although obviously high speed is a bonus) Connectivity you see is from a single mobile phone. A car could have a better antenna, service from all the mobile networks rather than just one, wifi fallback and a backup satellite uplink. If all those fail, you dispatch an engineer to find the car wherever it's stuck and add a solar powered LTE repeater. If you have a map of the whole service area and know where you started and have a camera looking out the window the whole time, there is no need for GPS really.
  - EdHominem 10y ago
    It might even be the sort of thing where you tap the video of the street ahead to tag the bag as passable and the car then slowly drives through it. No remote guidance per-se. Considering these are car companies doing this, true remote control is not a good idea. The industry response to the car hacks was pathetic. Only Tesla reacted properly. (And probably Google...)

iraklism 10y ago

I get that we have to be open with tech. After all most tech seemed "impossible" "unrealistic" and "unfeasible" even a couple of years before they go mainstream. But you have touched some use cases that are more frequent than most people think. London is a great example. My hometown Athens is much more closer to London as far as driving is concerned as opposed to LA or NY. I guess many other metropolis also. How will AI brand A communicate with AI brand B about life critical decisions ? Hey we can either both crash to each other or kill that kid on the pavement. ( just an example , I don't want to go to the morality of the situation) I assume there is going to be a "standard " . But have a look at your everyday tech. Standards are not easy. There are still WiFi routers/repeaters that cannot talk to each other nicely, just to give an example. Usually in life critical systems you don't have interoperability issues because there is no interoperability. Huge fail-safe systems from a single vendor.

walrus1066 10y ago
I can see communication with other autonomous cars happening, but with humans is another question altogether. We enter the realms of 'general ai' here, which is way out of scope of our ML tools.
- ThomPete 10y ago
  Is that really true though? You don't use your general AI to see someone extending their hand in traffic ex. General AI is the ability to both drive car, score a girl, cook food, clean your room etc. I think you would be surprised how little general AI have to do with this.
  - walrus1066 10y ago
    Sure, not Turing test level AI, but the car needs not to just correctly interpret what the other human driver is communicating, it too needs to communicate back, sometimes assertively. Sometimes you need to actually talk to the other driver. Anecdote: I was driving down a long, narrow country road, lots of turns, space only for 1 car, land rover coming in opposite direction, and another car behind me. We had to verbally negotiate who reverses back 50m to let the others go past. I cannot think how an ML algo could resolve this situation, even if it could, how would it communicate it's intentions to the humans, when flashing your lights isn't enough?

woofyman 10y ago

And that's just a partial list: deep pot holes, construction zones, standing water an inch deep or is it a foot deep?

lifeformed 10y ago
How do humans tell how deep water is?
- trome 10y ago
  Humans generally look at it, make an informed guess and drive accordingly. The recent Tesla accident is a great example of how even the best optics can be blown out by environmental factors like sun, whereas a human would likely readjust their vantage point, lower the sunshade, put on sunglasses or even pull over if they are unable to cope. Comparatively, the car did not attempt to do any of this and actually didn't even realize its cameras were in a state of failure.
  - gsam 10y ago
    Which says a lot about relying specifically on optics. Readjusting vantage points from an autonomous vehicle is as simple as having multiple sensors and deciding to rely more on a particular sensor in particular conditions. In particular, Google has some of the most expensive arrays of sensors.

brosky117 10y ago

The page about technology says, "Our sensors observe that the cyclist has extended their left arm. Detecting the cyclist's hand signal, our software predicts that the cyclist will move to the left side of the lane." This suggests to me that they're trying to deal with situations like the ones you've brought up.

moxious 10y ago
There is a difference between "trying to deal with" and "product on sale next month". These cars are overhyped. They've been just around the corner for the last 10 years. They are the Duke Nukem Forever of Silicon Valley.
- criley2 10y ago
  >These cars are overhyped Not really, they're correctly hyped. The concept is revolutionary. > They've been just around the corner for the last 10 year B-S. Holy revisionism BS!! We were NOT talking about self-driving cars in 2006! There was no company seriously designing fully autonomous cars in 2006! It was fully in the realm of science fiction, no investment, no major players, nothing. And you make it sound like it was supposed to be released 10 years ago! Vaporware! They told us we'd have AI cars in 2006! It never came! <---- False reality that never occurred. Why do you feel the need to lie about recent history? Shouldn't you observe that if you have to make a nasty lie to support your argument, maybe your argument is wrong? You're definitely parroting the contrarian point here, but you're being very low-effort about it. Contrarian can be cool but please try harder.
  - ZeroFries 10y ago
    https://en.wikipedia.org/wiki/History_of_autonomous_cars https://en.wikipedia.org/wiki/History_of_autonomous_cars
    - ssalazar 10y ago
      A link to a 15 page wikipedia article is not a rebuttal.

swamp40 10y ago

They don't need to make them foolproof. They just have to make them safer than human drivers. And these days with everyone swerving as they text, the bar isn't exactly set very high.

legolas2412 10y ago
They need to make it safer than average driver being adviced by driver assistance systems. That system can fix the negatives of human drivers without affecting the advantages of humans over machines

BurningFrog 10y ago

A lot of that will be best solved by making a lot of other actors communicate directly with the car. Traffic lights will send out the same messages they send to humans by red/green lights to the cars using radio. Emergency vehicles, roadworks, traffic signs etc will do the same. Or so I assume, because it makes so much of sense. I know there is vehicle-to-vehicle communication in the works: https://en.wikipedia.org/wiki/Vehicle-to-vehicle https://en.wikipedia.org/wiki/Vehicle-to-vehicle

walrus1066 10y ago
Yes, I think you're right. We'd need to update a lot of our infrastructure to make it friendlier for these cars. I should also add: - very fast mobile data, with no blackspots, so a remote driver can take over a car that's confused by a situation - a replacement for GPS, it's not reliable enough, the signal is far too weak - digital signal beacons embedded in lane markings, and edges of the road, and on traffic cones for roadworks
- trome 10y ago
  Why should local municipalities foot the bill for adding beacons to stop lights, etc? I get trying to make infrastructure friendly towards self driving cars, but the general public should not be footing the bill for someone's life choice to buy such a vehicle.
  - walrus1066 10y ago
    Because if we do resolve these issues, our roads will be much, much safer with autonomous cars. 2k people die on the roads each year in the UK, multiples more sustain life long injuries. I was in an accident, which could have been prevented were it not for the 1 second it takes for a human to react to a situation.
  - BurningFrog 10y ago
    This is important traffic infrastructure, just like visual traffic lights and asphalt. That said, it can be paid for my many actors.

Angostura 10y ago

A lot of these problems can be solved by simply making them fly. ... that's right, now you know what the driving force behind all those flying cars in sci-fi was. It's simply a way to separate the autonomous and old-style vehicles.

elmar 10y ago
you are right, driverless cars are so passe the future are self-flying vehicles.
jugchock 10y ago
This seems to have been said in jest but the main thing keeping flying cars in sci-fi and out of the real world is the human element. Most people have a hard enough time driving in two dimensions. Adding a third would be chaos without the aid of our robot overlords.
- dragonwriter 10y ago
  > This seems to have been said in jest but the main thing keeping flying cars in sci-fi and out of the real world is the human element. I thought it was that winged flight makes for crappy cars and wingless, vectored thrust flight is horribly fuel hungry. (Sci-fi versions are usually wingless and either have magic drive systems or just ignore the fuel issue.)

bhntr3 10y ago

My scuba instructor shared this piece of wisdom: "You don't have to swim faster than the shark. You just need to swim faster than your buddy." People get this stuff wrong all the time. Especially when they're drunk, tired or texting. These cars won't be perfect but I think pretty good and conservative under uncertainty will be better than human, drunk and overconfident.

ticviking 10y ago
That's a key point here. The cars don't have to be perfectly safe, just safer than the average driver.
- woofyman 10y ago
  It has be better than the best human driver or they'll get sued out of existance. You can't sell a product that you know will hurt or kill a certain number of people.
  - deadbunny 10y ago
    I'm pretty sure the car industry has been doing that for years.
    - woofyman 10y ago
      They get sued if they knowingly sell a defective product.
  - ageofwant 10y ago
    Sorry no. Cars kill thousands of people every year, and will continue to do so. Car companies are doing just fine. Power companies literally kill millions every year, a-ok. McDonalds, Coca Cola ? ditto.
    - woofyman 10y ago
      Drivers kill, not the cars. That's totally different from an autonomous car killing someone. Humans make mistakes. That's different from a defective product.

LinuxBender 10y ago

I agree with all of this. I would also add; that I will stay consider these vehicles to be unsafe until the manufacturers can prove to be beyond a shadow of doubt that I or the local computer are the ones piloting the vehicle. When a random person can remotely take control of a vehicle; whether it be the windshield wipers, or full control of the vehicle, then it is not safe to contain a power source in the vehicle near a public road. I don't want 4chan driving me into another vehicle.

andrewljohnson 10y ago
I'd consider autonomous cars to be more safe than a car I drive when statistically I'm less likely to get hurt or die. If there was some chance my car could get hacked, or other weird vulnerabilities that robocars had that traditional cars did not, I would still want to be in a robocar if my chance of dying was cut in half.
- lazaroclapp 10y ago
  Sure. But you are assuming a good way to estimate that chance. Say you base it on previous performance: in the last 5 years, twice as many people have died driving non-self-driving cars compared to those riding self-driving cars. Then, the very next day, a massive hack / cyberattack happens that causes 10 million fatal accidents...
- brazzledazzle 10y ago
  This is going to be one of those things where a lot of people have trouble quantifying the actual risk because the fear when things are out of your control is so powerful. Like flying.
  - LinuxBender 10y ago
    Not to mention, the live hacks of moving vehicles have already occurred multiple times. Thankfully, there are not yet enough skiddie tools and the bar to entry is just high enough that mass casualties have not yet occurred. Give it a few more months or maybe even another year before your car is just another IoT that can be controlled remotely. I do not believe it would be irrational to consider the risk of 100k cars driving into oncoming traffic at the same time. That is a lot more damage and chaos than a few little 747's could do. Am I incorrect? My only questions are 1) will the skiddie tools be on github first, or will someone have to buy/leak them and repost on github? 2) If a minor executes the scripts and lives are taken; and assuming you can determine who ran them, will they be tried as adults and will the verdict or punitive actions make up for the damages?

yohui 10y ago

https://www.wired.com/2016/12/google-self-driving-car-waymo https://www.wired.com/2016/12/google-self-driving-car-waymo > By early 2015, Fairfield thought they were getting close: The cars were clocking full days at Google’s testing grounds without needing human takeovers. They had driven 1.2 million miles on public roads and could pull over for emergency vehicles. They could detect and brake for squirrels, and read hand signals from cops and construction workers. They knew when to honk politely (Just making sure you see me here) and when to blare it (You’re about to slam into me!). Then the team spent the next year putting another 800,000 miles on the fleet’s collective odometer, to fine tune everything. And it gave 10,000 rides to employees and “guests.”

walrus1066 10y ago
The car-human interaction bit isn't from a quote, is there any evidence of this actually happening in the field? Would be great if there was a video or something.
- marwatk 10y ago
  You can see one example at the end of the video on their technology page where the cyclist can't make up his mind. It recognizes the left turn sign from the rider (you can see the sybmol change when he extends his arm). https://www.youtube.com/watch?v=fbWeKhAPMig https://www.youtube.com/watch?v=fbWeKhAPMig
  - StRoy 10y ago
    This kind of predictable reactions to specific human gestures.. I fear it will have great uses for certain type of crimes. What will you do if someone stops you with the purpose of assaulting _for eg, a robbery_ you and the car refuses to run the dude over because he made a gesture and is now in front of your car while his buddies are waiting aside? There are good reasons why things like these happened: > The law of some states, such as Louisiana, explicitly lists a killing in the course of defending oneself against forcible entry of an occupied motor vehicle as a justifiable homicide. While I'm interested in self driving cars that still feature manual controls you can override, like Tesla, I have no interest in a future controlled by Google's AI without my input.

agumonkey 10y ago

All these are old "problems", either solved or in progress. Something I've never seen adressed is how SDV fleets behave when change occurs. Are there chaotic oscillation patterns that emerge from AI trying similar avoiding behavior. There's research in massive multi agent behavior, but I've never seen it tested on the field with SDV makers actual models.

spuz 10y ago

I think a good way to think about the last 5% automation problem is to look at other examples where automation has replaced human work. In every case that I can think of, the work of the human was never 100% replaced. Automated check outs at the supermarket still require a number of attendants. An automated assembly line still needs factory workers. Automated rail systems are still controlled by humans in a control room. You can also look at how the environment has been adapted to better suit the new technology. For example to allow big heavy horseless carriages roll along roads we replaced dirt with tarmac (asphalt). Driverless cars will never (until we have general AI) be able to fully replace a human driver. However we don't need to get to 100% for the technology to become useful. It might be that driverless cars will be restricted to only driving on certain mapped roads. Or every car's movements are tracked in a central control room and if a problem is detected either by the car or the passenger it could be controlled remotely. Or let's say a GPS sensor breaks an actual human driver might be issued to come and rescue you. Lastly we might even accept a certain failure rate. Society is willing to tolerate failure of technology to a certain degree in cars and planes today. All we need is is for the technology to make us all better off on average for it to be successful.

teknoxjon 10y ago

I'd imagine all the GPS data could be loaded into the system beforehand and updated when there is a consistent signal. The same way cars now deal with GPS.

nialv7 10y ago

> They need to reliably deal with an almost infinite number of edge and corner cases, each quite different from the last. It doesn't need to be perfect. It just needs to be better than average human drivers. Sure there will be cases it can't handle. But the same is much more true for human.

dilemma 10y ago
Red herring. It doesn't need to be perfect; it needs to handle an infinite number of edge cases. Humans can, machines cannot.
- nialv7 10y ago
  Strictly speaking, it needs to handle an infinite number of edge cases _in a very limited amount of time_ Machines might not be able to, humans definitely can't.

sandworm101 10y ago

>> communication with other human drivers. I would shorten that to "other humans", to specifically include pedestrians. That 1/2-second eye contact between driver and human is key. Would anyone here dare walk in front of a car without that little nod from the driver? The other edge case, one that is never discussed, is how these things expect to deal with off-road driving. I regularly use car ferries. Getting on and off those, and negotiating the terminals, stumps many human drives.

Shivetya 10y ago

I am still waiting for the at night in full on rain demos on unmarked roads or such.

redwood 10y ago

What you're missing is "how do humans cope with these scenarios?" -- by either cautiously slowing down or by following those humans ahead or both. Machines can do the same