4 ms·

Not just that, getting self-driving approved with regulators is a major challenge. Also, to achieve level 5 we need not just algorithm improvement, our physical

Not just that, getting self-driving approved with regulators is a major challenge. Also, to achieve level 5 we need not just algorithm improvement, our physical infrastructure need to be improved as well. On intersection, two human drivers wave at each other to decide who goes first. How should the human driver communicate with the machine? What should the machine do when there is a dog standing in the middle of a street? Move on and let a human driver take care of it? Isn’t that now a bystander effect? These are just some examples I am not sure if textbook definition of level 5 will address. Going really extreme: in a car accident in a somewhat remote area, the driver might be able to get an injured passenger out if the passenger stuck. Now a fully autonomous car, there is only passenger. Yes, for sure this is extreme and perhaps likely a dramatic scene only, but it is one of those edge cases we can’t cover in a self-driving car.

JoshTriplett 9y ago
> How should the human driver communicate with the machine? In theory, a self-driving car could watch for hand-based turn signals, just like a human driver could, and take those into account. In practice, what someone signals is far less interesting than what they do, and all it does is provide hints for reaction time, which 1) a computer already wins at, and 2) a computer should always leave enough time for braking/stopping/etc anyway. A human driver should never make an assumption based on someone's turn signal that would lead to a crash if they don't do what they're signaling, and neither should a computer. > What should the machine do when there is a dog standing in the middle of a street? Don't hit it (stopping if necessary), and listen when the human in the vehicle says "stop" (or "call animal control", or any number of other reactions). The human in the vehicle should always be in control of the high-level actions, just not the moment-to-moment ones. > Going really extreme: in a car accident in a somewhat remote area, the driver might be able to get an injured passenger out if the passenger stuck. Now a fully autonomous car, there is only passenger. Not sure what scenario you're getting at here. If you mean a driver and passenger in the same vehicle, the same problem exists today with an injured driver.
philwelch 9y ago
The algorithm doesn’t have to be perfect, it just has to be a significant improvement over human beings. For every fatality an algorithm causes that a human would not, there could easily be ten fatalities prevented.
- QAPereo 9y ago
  You do realize that most people won’t accept that, right? Worse, when the algorithm does kill someone, who is responsible? Edit: Instead of just downvoting and moving on, why not actually formulate a response? Is there anything more empty than disagreement without the capacity to argue your case?
  - golfer 9y ago
    This is currently being tested in court. Tesla's autopilot has resulted in fatalities.
    - kelnos 9y ago
      That's different, though. Autopilot doesn't absolve the driver of the car from responsibility to be alert and take control when necessary, so the argument can be made that the responsibility is shared, or even rests on the human driver, depending on the circumstances. But a level 5 system will theoretically not even have a driver, and might not allow passengers to control the vehicle at all aside from giving it a destination. I think that becomes easier, though: in the case where no human in the car has control over what the car does, it's obviously the manufacturer's fault when something goes awry. Though if doing something like rapidly and repeatedly changing the car's destination and intentionally "confusing" it leads to a crash, that might shift blame to one of the passengers, which would be unusual.
  - Jedd 9y ago
    Won't accept which - that computers only have to beat human fatality rates to be preferable, or that the ratio will be 10:1 (or better) in favour of computer-driven cars in terms of fatalities? I'd agree that if we aren't seeing better fatality numbers, people won't accept it -- but there's little to suggest that humans are a difficult target to beat. Especially in the US, where car-related fatalities are so numerous and much of the research and trials are being done. Working out accountability for fatalities is already a hugely complicated process, involving an unhealthy number of law enforcement and legal personnel. The problem of identifying responsible party for a - as you put it - failed algorithm doesn't seem intractable, and certainly isn't a show-stopper. Haven't voted either way on this thread, but I suspect your re-hashing of regular topics already resolved is the problem.

JoshTriplett 9y ago

> How should the human driver communicate with the machine? In theory, a self-driving car could watch for hand-based turn signals, just like a human driver could, and take those into account. In practice, what someone signals is far less interesting than what they do, and all it does is provide hints for reaction time, which 1) a computer already wins at, and 2) a computer should always leave enough time for braking/stopping/etc anyway. A human driver should never make an assumption based on someone's turn signal that would lead to a crash if they don't do what they're signaling, and neither should a computer. > What should the machine do when there is a dog standing in the middle of a street? Don't hit it (stopping if necessary), and listen when the human in the vehicle says "stop" (or "call animal control", or any number of other reactions). The human in the vehicle should always be in control of the high-level actions, just not the moment-to-moment ones. > Going really extreme: in a car accident in a somewhat remote area, the driver might be able to get an injured passenger out if the passenger stuck. Now a fully autonomous car, there is only passenger. Not sure what scenario you're getting at here. If you mean a driver and passenger in the same vehicle, the same problem exists today with an injured driver.

philwelch 9y ago

The algorithm doesn’t have to be perfect, it just has to be a significant improvement over human beings. For every fatality an algorithm causes that a human would not, there could easily be ten fatalities prevented.

QAPereo 9y ago
You do realize that most people won’t accept that, right? Worse, when the algorithm does kill someone, who is responsible? Edit: Instead of just downvoting and moving on, why not actually formulate a response? Is there anything more empty than disagreement without the capacity to argue your case?
- golfer 9y ago
  This is currently being tested in court. Tesla's autopilot has resulted in fatalities.
  - kelnos 9y ago
    That's different, though. Autopilot doesn't absolve the driver of the car from responsibility to be alert and take control when necessary, so the argument can be made that the responsibility is shared, or even rests on the human driver, depending on the circumstances. But a level 5 system will theoretically not even have a driver, and might not allow passengers to control the vehicle at all aside from giving it a destination. I think that becomes easier, though: in the case where no human in the car has control over what the car does, it's obviously the manufacturer's fault when something goes awry. Though if doing something like rapidly and repeatedly changing the car's destination and intentionally "confusing" it leads to a crash, that might shift blame to one of the passengers, which would be unusual.
- Jedd 9y ago
  Won't accept which - that computers only have to beat human fatality rates to be preferable, or that the ratio will be 10:1 (or better) in favour of computer-driven cars in terms of fatalities? I'd agree that if we aren't seeing better fatality numbers, people won't accept it -- but there's little to suggest that humans are a difficult target to beat. Especially in the US, where car-related fatalities are so numerous and much of the research and trials are being done. Working out accountability for fatalities is already a hugely complicated process, involving an unhealthy number of law enforcement and legal personnel. The problem of identifying responsible party for a - as you put it - failed algorithm doesn't seem intractable, and certainly isn't a show-stopper. Haven't voted either way on this thread, but I suspect your re-hashing of regular topics already resolved is the problem.