4 ms·

I strongly suspect we will soon enough find out that Waymo hit a wall in self-driving capability possible with ML methods and got rid of its in-car safety drive

I strongly suspect we will soon enough find out that Waymo hit a wall in self-driving capability possible with ML methods and got rid of its in-car safety drivers only by moving the actual humans to a central facility where they monitor and assist the vehicles remotely. And that plus the extreme maintenance required to maintain the cars’ sensors is going to keep Waymo and any other such company from being able to roll out an economically viable taxi service at any meaningful scale.

mlyle 6y ago
I don't think so. I've watched a lot of Waymo videos and it's pretty clear when the car needs remote assistance help. Mostly parking lot situations. It's still likely a bit too much, but I think they don't need much further progress. The bigger problem is that they still have absolutely no confidence in any weather worse than when it's barely sprinkling... then the safety drivers come back out.
- SV_BubbleTime 6y ago
  There are so many weather scenarios that are obstacles we as humans barely overcome. I think about it like the meme ”Almost at my destination, better turn off my radio so I can see where I’m going”. That is we as humans the best drivers in the universe so far as we know have to limit some sensory intake so we can focus that every elsewhere to make even simple calculations. Asking a vehicle sensor suite to handle white out snow or fog with halos around lights or to ignore its lidar in super heavy rain just seems like the equivalent of blaring the radio when you’re trying to get across 8 lanes of superhighway traffic and it’s your first day in a left hand drive vehicle.
  - mlyle 6y ago
    Twenty years ago I handled situations in sensor fusion where some of the sensors go to shit in specific circumstances, and it wasn't blazing new ground. The state of the art and the tools to deal with this have grown massively since then.
    - SV_BubbleTime 6y ago
      I’m an automotive EE, so I encourage you to come fix all the problems. Because imo we’re 5 years out from where people think we are today. Maybe the problem isn’t practical, it’s just we just haven’t dumped enough brain power into it.
      - mlyle 6y ago
        I think there's massive problems, but the one you described -- attentional saturation -- is not the way sensing systems behave. That is a human issue. Trying to anthropomorphize the computer solving the problem is not a path to understanding the problem space.
        SV_BubbleTime 6y ago
        As a 1:1 exact problem, no of course. As an analog to challenges, sure. There is also the “who is more important” problem? A self-interst version of the trolly problem. When asked this, a Mercedes executive explained that the driver was the most important person to the vehicle. So given the option of hit this biker or swerve into oncoming traffic, they would chose the biker. So, somewhere in software there is a decision to always protect the driver, interesting problem that we have no good answer for. Back to my point though, I’m skeptical we’re even close to having all the data to make the decision properly.
        mlyle 6y ago
        > As a 1:1 exact problem, no of course. As an analog to challenges, sure. I disagree. > So given the option of hit this biker or swerve into oncoming traffic, they would chose the biker. You act like this is really complicated, but given that self-driving cars generally are quite conservative in their dynamics and where they choose to go, this is pretty much the default behavior. Swerving into oncoming traffic is just not in their DNA right now, even in the rare circumstances where it's advantageous. Right now there's some work on aggressive dynamics for difficult situations, but this will tend to be underused not overused (just like it is by humans). Doing the default, conservative thing is generally the self-protective one.
jahnu 6y ago
If autonomous driving is approaching a sophistication that is practically useful on a large scale shouldn't we first see it widely deployed for delivery of packages and food and the like before we see it being used to move much more valuable charge like people?
- ghaff 6y ago
  They'd still be on the road with other people and the actual delivery of a package requires quite a bit of "common sense" absent some sort of drop-off zone with a beacon.
  - jahnu 6y ago
    I agree but some problems disappear too. For example, occupants of cars get frustrated when they are stuck for quite some time on a complicated turn that a human would accomplish quickly. Packages don't care either if the driving is not smooth and the vehicle responds in favour and with extra care of other vehicles around it. Not to mention routes that might pick up and deliver multiple simultaneous loads. It just seems that would be an easier first step for the tech that should be profitable. But these are just questions I have. I'm very unsure about what the reality is.
- redisman 6y ago
  Deliveries are made to people’s doorsteps or a lobby/mailroom at least in the US. Not sure how a car could do that. Maybe with a robot throwing arm?
  - _flux 6y ago
    Actual drones, of course!
  - IshKebab 6y ago
    People would come out to the car. There are already small experimental food delivery robots.
- IshKebab 6y ago
msoad 6y ago
If problem was cost of building and maintaining hardware they would go after that problem by iterating on hardware much faster. That is not the immediate problem. What they need to solve first is to reach to a level of confidence with "happy ride ratio" in which they don't need to hire too many remote assistant drivers to scale up. After all, this technology wants to remove human drivers from the taxi business. What's the point of paying often more expensive remote human drivers plus super expensive gear to do a low margin business like taxi service?
- lupire 6y ago
  It's a lot worse than that. The reason Waymon is in Arizona is that it has almost no weather. Remote assistance isn't enough to rescue a ride in weather. Uber bought out almost the entire CS department of CMU (Pittsburgh) yet they run their cars in Arizona.
  - mlyle 6y ago
    It's kinda an interesting problem-- the technology readiness re: weather. Right now a lot of self driving, like Waymo, is aimed at taxi fleets, because the equipment is expensive, so amortizing the cost over many driving hours per day makes fiscal sense. But-- a taxi fleet that requires a massive scale up / instant appearance of a human workforce when it's rainy is a non-starter. On the other hand, personal cars would be useful with an "only sunny" autonomous mode.
    - ghaff 6y ago
      To a discussion upthread, it sometimes seems as if many of the folks working on full autonomy decided to deliberately turn their difficulty up to 11. What use case shall we tackle first? Highway driving in good weather? "No that's boring. How about largely-urban taxis in all weather types?"
      - mlyle 6y ago
        The problem is, it's not clear that camera-only and relatively modest processing can even handle fully autonomous highway driving in good weather. So, if you need a very expensive sensor package, the only large market available is taxis.

mlyle 6y ago

I don't think so. I've watched a lot of Waymo videos and it's pretty clear when the car needs remote assistance help. Mostly parking lot situations. It's still likely a bit too much, but I think they don't need much further progress. The bigger problem is that they still have absolutely no confidence in any weather worse than when it's barely sprinkling... then the safety drivers come back out.

SV_BubbleTime 6y ago
There are so many weather scenarios that are obstacles we as humans barely overcome. I think about it like the meme ”Almost at my destination, better turn off my radio so I can see where I’m going”. That is we as humans the best drivers in the universe so far as we know have to limit some sensory intake so we can focus that every elsewhere to make even simple calculations. Asking a vehicle sensor suite to handle white out snow or fog with halos around lights or to ignore its lidar in super heavy rain just seems like the equivalent of blaring the radio when you’re trying to get across 8 lanes of superhighway traffic and it’s your first day in a left hand drive vehicle.
- mlyle 6y ago
  Twenty years ago I handled situations in sensor fusion where some of the sensors go to shit in specific circumstances, and it wasn't blazing new ground. The state of the art and the tools to deal with this have grown massively since then.
  - SV_BubbleTime 6y ago
    I’m an automotive EE, so I encourage you to come fix all the problems. Because imo we’re 5 years out from where people think we are today. Maybe the problem isn’t practical, it’s just we just haven’t dumped enough brain power into it.
    - mlyle 6y ago
      I think there's massive problems, but the one you described -- attentional saturation -- is not the way sensing systems behave. That is a human issue. Trying to anthropomorphize the computer solving the problem is not a path to understanding the problem space.
      - SV_BubbleTime 6y ago
        As a 1:1 exact problem, no of course. As an analog to challenges, sure. There is also the “who is more important” problem? A self-interst version of the trolly problem. When asked this, a Mercedes executive explained that the driver was the most important person to the vehicle. So given the option of hit this biker or swerve into oncoming traffic, they would chose the biker. So, somewhere in software there is a decision to always protect the driver, interesting problem that we have no good answer for. Back to my point though, I’m skeptical we’re even close to having all the data to make the decision properly.
        mlyle 6y ago
        > As a 1:1 exact problem, no of course. As an analog to challenges, sure. I disagree. > So given the option of hit this biker or swerve into oncoming traffic, they would chose the biker. You act like this is really complicated, but given that self-driving cars generally are quite conservative in their dynamics and where they choose to go, this is pretty much the default behavior. Swerving into oncoming traffic is just not in their DNA right now, even in the rare circumstances where it's advantageous. Right now there's some work on aggressive dynamics for difficult situations, but this will tend to be underused not overused (just like it is by humans). Doing the default, conservative thing is generally the self-protective one.

jahnu 6y ago

If autonomous driving is approaching a sophistication that is practically useful on a large scale shouldn't we first see it widely deployed for delivery of packages and food and the like before we see it being used to move much more valuable charge like people?

ghaff 6y ago
They'd still be on the road with other people and the actual delivery of a package requires quite a bit of "common sense" absent some sort of drop-off zone with a beacon.
- jahnu 6y ago
  I agree but some problems disappear too. For example, occupants of cars get frustrated when they are stuck for quite some time on a complicated turn that a human would accomplish quickly. Packages don't care either if the driving is not smooth and the vehicle responds in favour and with extra care of other vehicles around it. Not to mention routes that might pick up and deliver multiple simultaneous loads. It just seems that would be an easier first step for the tech that should be profitable. But these are just questions I have. I'm very unsure about what the reality is.
redisman 6y ago
Deliveries are made to people’s doorsteps or a lobby/mailroom at least in the US. Not sure how a car could do that. Maybe with a robot throwing arm?
- _flux 6y ago
  Actual drones, of course!
- IshKebab 6y ago
  People would come out to the car. There are already small experimental food delivery robots.
IshKebab 6y ago

msoad 6y ago

If problem was cost of building and maintaining hardware they would go after that problem by iterating on hardware much faster. That is not the immediate problem. What they need to solve first is to reach to a level of confidence with "happy ride ratio" in which they don't need to hire too many remote assistant drivers to scale up. After all, this technology wants to remove human drivers from the taxi business. What's the point of paying often more expensive remote human drivers plus super expensive gear to do a low margin business like taxi service?

lupire 6y ago
It's a lot worse than that. The reason Waymon is in Arizona is that it has almost no weather. Remote assistance isn't enough to rescue a ride in weather. Uber bought out almost the entire CS department of CMU (Pittsburgh) yet they run their cars in Arizona.
- mlyle 6y ago
  It's kinda an interesting problem-- the technology readiness re: weather. Right now a lot of self driving, like Waymo, is aimed at taxi fleets, because the equipment is expensive, so amortizing the cost over many driving hours per day makes fiscal sense. But-- a taxi fleet that requires a massive scale up / instant appearance of a human workforce when it's rainy is a non-starter. On the other hand, personal cars would be useful with an "only sunny" autonomous mode.
  - ghaff 6y ago
    To a discussion upthread, it sometimes seems as if many of the folks working on full autonomy decided to deliberately turn their difficulty up to 11. What use case shall we tackle first? Highway driving in good weather? "No that's boring. How about largely-urban taxis in all weather types?"
    - mlyle 6y ago
      The problem is, it's not clear that camera-only and relatively modest processing can even handle fully autonomous highway driving in good weather. So, if you need a very expensive sensor package, the only large market available is taxis.