3 ms·

> Once you see a car, it's fairly easy to make the algorithm pessimistically assume that any car that can cut it off is likely to do so. Well, this just doesn'

> Once you see a car, it's fairly easy to make the algorithm pessimistically assume that any car that can cut it off is likely to do so. Well, this just doesn't match reality. Uber showed us that these algorithms are full of false positives and have to be disregarded to make the vehicle actually drive itself (Remember when they killed that pedestrian?). What a computer would need is 'awareness,' and that awareness is even more critical in a larger body vehicle. There's always going to be 'emergency stop in case of unknown situation.' The difference is, the human is not in an unknown situation, they're able to anticipate and plan. Given the current landscape, I'd say with enough data you could more or less brute-force an AI driven car and be somewhere on the spectrum of safer than the worst human drivers. I say it's practically impossible for a semi truck to brute-force with sensors and pessimistic algorithms alone.

0xffff2 7y ago
Yes, I do remember when Uber killed a pedestrian. A pedestrian that the car didn't see. Yes the computer needs awareness, but it needs awareness no matter the kind of vehicle it is controlling. The hard part of awareness is the ability to see and correctly categorize potential hazards. That part is absolutely no different for any motor vehicle. The way the vehicle reacts to potential hazards is absolutely going to change depending on the vehicle's characteristics, but that's one of the least hard parts of the whole system. >Given the current landscape, I'd say with enough data you could more or less brute-force an AI driven car and be somewhere on the spectrum of safer than the worst human drivers. What's your definition of "enough data"? Tesla seems about as close are you could practically get, and they don't have a great safety history (although they don't have enough history period to really say whether they're safer than humans). The specific issues that Tesla and Uber have faced are issues with seeing and correctly interpreting the world around them. Those issues are largely the same in either case. If you assume you can see and follow lane markings and it turns out you don't do so well when the lane markings are damaged, you're going to have a bad time no matter what size vehicle you're in.
- linuxftw 7y ago
  > Yes, I do remember when Uber killed a pedestrian. A pedestrian that the car didn't see. The car did see the pedestrian. It was unable to correctly identify it as pedestrian initially. And instead of taking the pessimistic path of 'something unknown moving toward roadway, so I brake now' it had that portion disabled in favor of 'safety driver take over.' Reason being, as it came out, that these types of events are so common it would be wholly infeasible to attempt a 'driverless' solution without disabling the emergency braking routine. > lane markings are damaged, you're going to have a bad time no matter what size vehicle you're in. That's true. But it's not a problem humans have. In a Tesla (or whatever, not picking on them specifically), you can probably stop in a reasonable distance, and if you don't, you probably aren't going to kill scores of people. With a semi-truck, stopping distance is greatly increased. So, you need the ability to see farther up the road. If there's traffic in front of you, you obviously can't see lane markings 200 yards up the street. There's just so many variables. What if the road's wet and the lane markings end and the computer thinks it needs to slam on the brakes? I guess jack-knifing the trailer in the middle of busy highway is NBD, because that's what the computer was supposed to do. I won't share the road with such a contraption.

0xffff2 7y ago

Yes, I do remember when Uber killed a pedestrian. A pedestrian that the car didn't see. Yes the computer needs awareness, but it needs awareness no matter the kind of vehicle it is controlling. The hard part of awareness is the ability to see and correctly categorize potential hazards. That part is absolutely no different for any motor vehicle. The way the vehicle reacts to potential hazards is absolutely going to change depending on the vehicle's characteristics, but that's one of the least hard parts of the whole system. >Given the current landscape, I'd say with enough data you could more or less brute-force an AI driven car and be somewhere on the spectrum of safer than the worst human drivers. What's your definition of "enough data"? Tesla seems about as close are you could practically get, and they don't have a great safety history (although they don't have enough history period to really say whether they're safer than humans). The specific issues that Tesla and Uber have faced are issues with seeing and correctly interpreting the world around them. Those issues are largely the same in either case. If you assume you can see and follow lane markings and it turns out you don't do so well when the lane markings are damaged, you're going to have a bad time no matter what size vehicle you're in.

linuxftw 7y ago
> Yes, I do remember when Uber killed a pedestrian. A pedestrian that the car didn't see. The car did see the pedestrian. It was unable to correctly identify it as pedestrian initially. And instead of taking the pessimistic path of 'something unknown moving toward roadway, so I brake now' it had that portion disabled in favor of 'safety driver take over.' Reason being, as it came out, that these types of events are so common it would be wholly infeasible to attempt a 'driverless' solution without disabling the emergency braking routine. > lane markings are damaged, you're going to have a bad time no matter what size vehicle you're in. That's true. But it's not a problem humans have. In a Tesla (or whatever, not picking on them specifically), you can probably stop in a reasonable distance, and if you don't, you probably aren't going to kill scores of people. With a semi-truck, stopping distance is greatly increased. So, you need the ability to see farther up the road. If there's traffic in front of you, you obviously can't see lane markings 200 yards up the street. There's just so many variables. What if the road's wet and the lane markings end and the computer thinks it needs to slam on the brakes? I guess jack-knifing the trailer in the middle of busy highway is NBD, because that's what the computer was supposed to do. I won't share the road with such a contraption.