8 ms·

It's those edge cases that make me think real FSD (vehicles with no steering wheels) is a decade-timescale problem. Figure out 99.9% of driving, but otherwise

• alpha_squared 5y ago
  Just like reliability SLAs, I suspect every extra nine requires an exponential increase in investment. My armchair-know-nothing assessment is we're maybe at 90% coverage* right now, with maybe 10x all investment up until this point needed to reach 99%; then 10x that to reach 99.9%. Even at 99.9%, though, a failure scenario still threatens injuring or killing 0.1% of drivers (350,000 motorists for US alone). * Coverage as a percentage of scenario's occurrence over the total duration of driving. For example, over 90% of a long-distance trip will be spent on a highway following traffic patterns within a lane with the occasional lane change.
  • walshemj 5y ago
    Would not they detect the gates where down ? or are there still that many crossings without gates in the USA
    • ecpottinger 5y ago
      The place I talked about had no gates, after the accident they added gates.
    • bentcorner 5y ago
      Gates are only a suggestion - signals/gates malfunction so often you can find compilation videos of them on youtube.
    • rurp 5y ago
      Where I live in the Southwest US there are many crossings without gates. I'd say very few of the ones outside of well populated areas have them.
  • maxerickson 5y ago
    We regularize traffic more than your estimate though. Like 99.99% of the 300 mile drive I took yesterday was sitting in a lane. The hardest part was that Google hadn't mapped a service drive, so it thought the adjacent service drive was the best route (which would be addressed pretty fast if you were trying to deploy self driving service in that area). I don't think we get to level 5 very soon, but level 4 cars will have the ability to go lots of places pretty soon. If I overestimate driveway distances from yesterday, it's like 99.966% lane miles. There was some construction, but it was already well marked for human drivers.
• okareaman 5y ago
  I used to live in Mountain View where Waymo tests their cars so I would observe them quite often. I sat a railroad crossing with one, thinking about what could go wrong. What if the signal didn't work, but you could hear the train coming and see it way off to the right or left if you looked. Are Waymo cars listening for trains? Or are they looking at railroad guards only? I suppose they could program the train schedule in, but what if the train system changes the schedule? That could be adjusted for, but how much would I trust this system? Not enough to take a nap behind the wheel.
  • jeffbee 5y ago
    This seems like the edgiest of all possible edge case arguments against self-driving cars, considering that a meat-driven car crashed through a working barrier and into a perfectly obvious train in the Bay Area just this morning.
    • mulmen 5y ago
      Obviously meat-based drivers have flaws. The question is if the FSD robots at least retain all the existing capabilities of meat-based drivers. Trading one set of deficiencies for another raises the question of which set is preferable. The reality is that for a long time we will combine both sets of capabilities and use "self driving" tech to enhance human driver capabilities. In that case self-driving first needs to be able to avoid the relatively simple case of not of smashing through a barrier and the human driver can use their wetware to figure out how to handle railroad crossings which are diverse and complicated.
    • okareaman 5y ago
      This is where you're wrong because I'm obviously a far better driver than that person who drove in front of a train this morning and it would never happen to me. I trust myself more than I trust other drivers. I trust myself more than I trust Waymo. I'd be happy to have Waymo prove me wrong.
    • Swenrekcah 5y ago
      I agree that it’s an edge case but there are always completely bonkers people, out of the world on drugs or just suicidal. I don’t want to risk getting figuratively in a car with one if I turn on FSD.
• echelon 5y ago
  > It's those edge cases that make me think real FSD (vehicles with no steering wheels) is a decade-timescale problem. I'd be willing to bet it's a quarter century or longer problem. (Longbets, anyone?) FSD is absolutely achievable, but the task is much bigger than some proponents give it credit.
• ErikVandeWater 5y ago
  It is mostly a hardware problem. Tesla's can't do FSD with current hardware because there aren't redundant cameras in each direction. One camera gets mud on it and your car runs itself into oncoming traffic. (It does that anyway right now, but even if software bugs are fixed this is still intractable). A real solution has to be capable of cleaning all cameras (probably faster than a windshield wiper would), because the distortions caused by even normal rain are hellish to train an AI to handle.
  • xnx 5y ago
    Tesla would be lucky if the extent of its challenges were having enough cameras pointing in the right direction. Tesla handicapped itself by only developing its driving systems with cameras and not lidar.
    • ecpottinger 5y ago
      Right, because people have LIDAR installed already.
    • ErikVandeWater 5y ago
      I don't disagree. But since adding more cameras to a car that already has cameras is a much more intuitive extension for Tesla to select, I used that as an example.
• grecy 5y ago
  > but otherwise take a family off a bridge when the sun is blinding the camera? I love these wildly over-the-top exaggerations. When was the last bridge you saw without a barrier to prevent going off the edge? What makes you think a vehicle vision system will handle "blinded by sun" any worse than humans already do? Remember it's projecting and predicting the road ahead, even around corners and in the dark - so being blinded by the sun isn't going to cause it to swerve wildly off course and off the bridge - it can continue to use the data it had before being blinded (just as you do). Also remember it has eight cameras it uses for this. The 16 year old new driver texting and talking to friends coming towards you at 60mph has two.
  • xnx 5y ago
    "Blinded by the sun" is an understandable failure-state when you design a driving system that's reliant on cameras. From https://arstechnica.com/cars/2019/05/feds-autopilot-was-active-during-deadly-march-tesla-crash/ https://arstechnica.com/cars/2019/05/feds-autopilot-was-acti...: "Theoretically, it should be possible to detect the side of a truck using cameras. But it's not always easy. In some lighting conditions, for example, the side of a trailer might be hard to distinguish from the sky behind it."
    • grecy 5y ago
      How many of the 100 daily fatalities on the roads in the US do you think are caused by drivers not "seeing clearly"? The goal of an AI driver isn't no crashes. The goal is less crashes than human drivers.
      • tayo42 5y ago
        Idk if it's that simple, what if you reduce highway deaths by 100 but rural roads go up by 10. If I'm primarily a rural driver that's not enticing.
        • grecy 5y ago
          OK then, let's not put it into full scale production until all types of driving are an order of magnitude (10x) less.
• theluketaylor 5y ago
  The edge case that I always go to on why truly autonomous vehicles are decades+ away is winter driving. Lane markers and road signs disappear. Slush and salt spray constantly obscure sensors. Winter driving as a human driver requires an entirely different approach and Waymo hangs out in the Arizona desert where there is basically never any inclement weather.
  • astrange 5y ago
    You can simulate poor weather with sensors by either 1. simulating it or 2. driving at night.
  • heavyset_go 5y ago
    That, and construction, detours and city driving. In the latter, driving contexts can change without a second's notice, without any signs, markings or signals denoting the change. You learn what those contexts are from experience and understanding of what are ultimately complex social situations, and a shared understanding between you and the people around you.
  • Barrin92 5y ago
    don't even need to go to bad environmental setting. Let's talk about other countries. I haven't seen a lot of self-driving cars in Rome or Mumbai yet, there's a lot of places where traffic rules are treated more like suggestions than actual rules. American style neat wide highways and grid-like streets are not how the places look where most cars are being driven right now already.
    • sumedh 5y ago
      Mumbai literally has bumper to bumper traffic, I think any self driving car's sensors would just go crazy and give up.
  • klmadfejno 5y ago
    Digital data instead of signs seems like the obvious solution.
  • 5y ago
• mikepurvis 5y ago
  Or you need a high-bandwidth coverage-everywhere comms system and a robust infrastructure for remote teleop, like what Starsky was building: https://medium.com/starsky-robotics-blog/the-end-of-starsky-robotics-acb8a6a8a5f5#40fe https://medium.com/starsky-robotics-blog/the-end-of-starsky-... Then the steering-wheel-less car at least has the ability to call for help when it's only 98% sure of what to do instead of 99.99% sure. But obviously this kind of model only makes sense in a fleet context, not as an extension to something you own, so it requires greater shifts, at least for personal automobiles.
  • kgin 5y ago
    Most remote assistance is not live driving. A remote human adds labels and maybe even marks out a safe path for the vehicle so that vehicle has enough information to proceed.
    • mikepurvis 5y ago
      Starsky's was. Did you see the picture of their console?
• Spooky23 5y ago
  I think you’re right. For the AI we have today, it’s the equivalent of 1990. My guess is that as sensor arrays improve, wireless networks improve and quantum laptops are available in 2040 or so. Think about it — most software is deployed cloud first these days, but one of the most complex computing tasks we have is relying on some black box computer.
• nostrademons 5y ago
  IMHO FSD cars (along with drones, secure programming languages/OSes, and distributed manufacturing) are a post-war technology. The economics don't make sense while we're at peace and locked into the Pax Americana economy. However, once folks start trying to kill each other, these technologies will give a very large survival advantage to those factions that adopt them. Imagine how useful not having to risk human lives on supply lines would be. And risk expectations get reset when things go from "nobody ever dies unnaturally" to "people are actively trying to kill you", which eliminates the biggest barrier to FSD adoption. I think we'll get such a war within the next decade, so we may see FSD vehicles sooner than expected, just not in the way we want.
  • usrusr 5y ago
    In a hot industrialized war (one that isn't wildly asymmetric), a supply line has so many soldiers' lives depending on it that having the transports manned or not is insignificant. But I think I know the general concept you are getting at, I'd call it the wartime risk economy: when a large percentage of your combat aircraft sorties don't return, it becomes reasonable to operate the engines much closer to the limits of reliable operation than in peace time. The extra performance can save more pilots than the occasional engine failures caused by the emergency operation mode cost.
• Wowfunhappy 5y ago
  That shouldn't be the standard. It probably will be, but it shouldn't be. Humans are terrible drivers. If a self-driving car got into half as many accidents as the average human, it would save millions of lives. And kill people, to be sure, but fewer on net. I also think you could make a reasonable argument that all cars should be banned, right now, based on how many people they kill, but since I don't think that's gonna happen...
  • markkanof 5y ago
    You are correct on the aggregate, but that doesn't necessarily work when applied to each individual driver. I'm a pretty attentive and cautious driver. In the 20 years I have been driving I've been in one accident and that was because another driver was attempting to make an illegal left turn, came across two lanes of traffic, and t-boned me. So if self driving vehicles are only doing better than the worst human drivers, I'm going to be pretty hesitant to turn over control. I'd be in favor of that other driver that hit me using an autonomous vehicle though.
    • Wowfunhappy 5y ago
      Everyone thinks they're an above-average driver. Half of them must be wrong, so the reliability of any individual's self-assessment is about equal to chance. And even if you're one of the genuinely good drivers—your record certainly sounds excellent—our laws should account for the fact that most drivers can't judge their own skill.
      • mattkrause 5y ago
        ...and they might be right. It doesn't seem impossible for there to be a long tail of driving (in)ability: most people drive pretty well, with a small fraction that are distractible or reckless enough to account for most accidents.
  • usrusr 5y ago
    Just wait until self-driving "personality" becomes part of a car company's brand identity. I suspect that we will see no lack cars that are designed to be terrible drivers. Of course they will call it "confident", "assertive", perhaps for some brands even "masculine". Cars could become even more civilian tanks thank that already are. I desperately hope that I'm wrong.
• fnord77 5y ago
  Driving in all conditions requires the computing power of a human brain. When we have computers that powerful, we can talk about L5.
  • pcwalton 5y ago
    If that logic were valid, then Garry Kasparov would have beaten Deep Blue in 1997.
    • fnord77 5y ago
      chess doesn't require the processing of rapidly changing visual, inertial, tactile and auditory information while at the same time performing fine motor-coordination.