4 ms·

> The Google cars have been rear-ended seven times, often when stopped "but also on the freeway," wrote Urmson, director of Google's self-driving car program. I

by danhak 11y ago

> The Google cars have been rear-ended seven times, often when stopped "but also on the freeway," wrote Urmson, director of Google's self-driving car program. In other collisions, the cars were side-swiped or "hit by a car rolling through a stop sign." Eight of the 11 collisions were on city streets. http://bigstory.ap.org/article/297ef1bfb75847de95d856fb08dc0687/ap-exclusive-self-driving-cars-getting-dinged-california http://bigstory.ap.org/article/297ef1bfb75847de95d856fb08dc0...

prawn 11y ago
Mentioning accidents makes them sound dangerous. Mentioning that they haven't been at fault in any of them changes that significantly. In their last monthly report they described the two incidents in detail and they were the fault of the human driver behind them and about as trivial as collisions get.
- hvidgaard 11y ago
  They're still worth mentioning though. A good human driver can anticipate those situations and minimize the risk. Eg. I will not hard break for a yellow if the driver behind me is tailgating me, I'm super aware of drivers coming out from side roads because some of them just assume that there isn't coming any cars because there usually isn't. Not that an AI cannot learn that, but they're not good in unknown situations.
  - prawn 11y ago
    I'm willing to bet that those would be considered very basic and obvious problems to the people developing these cars. They work on this stuff every hour of every day. The AI will know (and share with other cars) tendencies for individual intersections and eventually share broader traffic knowledge. There will come a time where we just cannot compete with that. They will know every intersection's peculiarities and we'll know only those that we frequent, at best. They'll adjust for weather conditions too where we'll go by gut feel - e.g., at intersection X, what did people tend to do in light rain, driving West in the morning. What about on a hot day going N-S in peak hour. They will have data at hand for every location, condition and time if need be.
    - hvidgaard 11y ago
      Of course it's obvious to the researchers developing the cars. That doesn't make it simple or easy to solve, something as basic and simple as the halting problem cannot be solved. Data does not help you much in those 0.001% of situations where there is no data or too small a dataset to conclude anything. An AI can only act according to it's training. If we for a moment assume that it comes from the collective experience of all self driving cars, it's only as good as it's sensors allow it to be. They are simply not there yet. Could an modern AI provide autonoumous commute on the highway today - sure. Can it be trusted to navigate from A to B in urban areas and countryside, most of the time. But "most of the time" simply isn't good enough.
      - prawn 11y ago
        What's hard about detecting the signal ahead as well as the driver behind? The car might just elect to push through the orange rather than stopping abruptly. It could check and make that decision faster than any of us. Or, if paused in traffic with a car approaching too quickly from behind, it could eat up some of the space in front of it to escape/minimise collision. It could balance this decision with the risk of starting a chain reaction by taking that course of action. No one's pretending that these vehicles are ready for prime time now, but that they're ahead of what many would expect as the schedule, and could quickly handle what you've listed as a human advantage.
        hvidgaard 11y ago
        Until it can with certaincy detect the child on the bike behind bushes heading towards the road, it's not ready for 100% autonomous driving. It really doesn't matter how smart or how much experience it does have, if the sensors are inferior to humans.
        prawn 11y ago
        For some time now they've been able to detect temporary road signs from amongst advertising and other general signage. i.e., changes in speed, maintenance warnings, etc. If a human can see the child on the bike behind bushes, the car can too. Plus it can do it from multiple cameras at once. And calculate the speed and trajectory. And quickly work out the optimal course of action. I really think you're overestimating the ability of humans (who get drunk or distracted, routinely speed, etc) and underestimating the progress of self-driving cars.

by danhak 11y ago

prawn 11y ago

Mentioning accidents makes them sound dangerous. Mentioning that they haven't been at fault in any of them changes that significantly. In their last monthly report they described the two incidents in detail and they were the fault of the human driver behind them and about as trivial as collisions get.

hvidgaard 11y ago
They're still worth mentioning though. A good human driver can anticipate those situations and minimize the risk. Eg. I will not hard break for a yellow if the driver behind me is tailgating me, I'm super aware of drivers coming out from side roads because some of them just assume that there isn't coming any cars because there usually isn't. Not that an AI cannot learn that, but they're not good in unknown situations.
- prawn 11y ago
  I'm willing to bet that those would be considered very basic and obvious problems to the people developing these cars. They work on this stuff every hour of every day. The AI will know (and share with other cars) tendencies for individual intersections and eventually share broader traffic knowledge. There will come a time where we just cannot compete with that. They will know every intersection's peculiarities and we'll know only those that we frequent, at best. They'll adjust for weather conditions too where we'll go by gut feel - e.g., at intersection X, what did people tend to do in light rain, driving West in the morning. What about on a hot day going N-S in peak hour. They will have data at hand for every location, condition and time if need be.
  - hvidgaard 11y ago
    Of course it's obvious to the researchers developing the cars. That doesn't make it simple or easy to solve, something as basic and simple as the halting problem cannot be solved. Data does not help you much in those 0.001% of situations where there is no data or too small a dataset to conclude anything. An AI can only act according to it's training. If we for a moment assume that it comes from the collective experience of all self driving cars, it's only as good as it's sensors allow it to be. They are simply not there yet. Could an modern AI provide autonoumous commute on the highway today - sure. Can it be trusted to navigate from A to B in urban areas and countryside, most of the time. But "most of the time" simply isn't good enough.
    - prawn 11y ago
      What's hard about detecting the signal ahead as well as the driver behind? The car might just elect to push through the orange rather than stopping abruptly. It could check and make that decision faster than any of us. Or, if paused in traffic with a car approaching too quickly from behind, it could eat up some of the space in front of it to escape/minimise collision. It could balance this decision with the risk of starting a chain reaction by taking that course of action. No one's pretending that these vehicles are ready for prime time now, but that they're ahead of what many would expect as the schedule, and could quickly handle what you've listed as a human advantage.
      - hvidgaard 11y ago
        Until it can with certaincy detect the child on the bike behind bushes heading towards the road, it's not ready for 100% autonomous driving. It really doesn't matter how smart or how much experience it does have, if the sensors are inferior to humans.
        prawn 11y ago
        For some time now they've been able to detect temporary road signs from amongst advertising and other general signage. i.e., changes in speed, maintenance warnings, etc. If a human can see the child on the bike behind bushes, the car can too. Plus it can do it from multiple cameras at once. And calculate the speed and trajectory. And quickly work out the optimal course of action. I really think you're overestimating the ability of humans (who get drunk or distracted, routinely speed, etc) and underestimating the progress of self-driving cars.