4 ms·

Maybe now is the time to start raising the bar on human and upcoming automated drivers. Both could have to pass a driving test that includes challenging, diffi

Maybe now is the time to start raising the bar on human and upcoming automated drivers. Both could have to pass a driving test that includes challenging, difficult situations. Day, night, rain, snow, fog, heavy traffic, missing lane lines, simulated pedestrians (including small children) leaping out between parked cars, and more. For further difficulty, both kinds of drivers could have to pass this within a x% margin of the fastest safe speed at which the maneuvers can be completed; no cheating by slowing down to traffic-disrupting speeds.

beambot 5y ago
I suspect that a challenging driver assessment between AVs and humans would be the most rapid driver of AVs possible... Just remember: Most drivers think they're above average. Curiously, autonomous car technology also makes it feasible for us to rate drivers -- it's like having a professional watching over your shoulder all the time. IIRC, Tesla used something similar to determine rollout of new FSD capabilities.
nullc 5y ago
The reason to hold humans and self driving cars to a different standard is that humans have already "class qualified" under ... well ... every condition that we drive in... with on the order of a billion driving years of total experience to the limitations of human-driving and far in excess of 10 million years of additional driving added per year along with a considerable amount of study across many decades. Though there are systematic flaws in human drivers they're reasonably well known given the level of experience we have, there are many mitigations in place in both our cars and road designs, and many of these flaws are intuitively modeled by the other humans on the road as well as pedestrians allowing for real-time compensation. The driverless cars, on the other hand-- we have radically less experience with as a class and there are many reasons to believe that vendor to vendor (e.g. due to sensor modalities) or even version to version will behave less self-consistently than human drivers do as a class. We know of some systematic faults in their operation which cause them to behave in ways that are highly surprising to humans, and we should expect to find many more as we gain experience. To the extent that the properties of self driving cars have been formally studied at all it's been primarily by their creators which are obviously self interested. Some of the practices used in early supervised self-driving also actively undermined understanding their safety (e.g. executing a hand over to the human once the car was already in an nonrecoverable unsafe state and then attributing the accident to human driving since at the time of the collision the human was back in charge, but maybe only by a hundred milliseconds; or companies committing perjury to use the DMCA to force down unfavorable videos created by others) This isn't an argument that we might not benefit from more rigorous driver testing or even from testing ideas that originated from self-driving car testing... but parity between driverless car tests and human testing shouldn't be a goal, particularly not when so many orders of magnitude separate our human driving experience from driverless experience.
akira2501 5y ago
There's not a lot of evidence to suggest this would have a major impact. The majority of fatal crashes involve drugs and alcohol and not people with "untested skills in challenging situations." Pedestrians and motorcyclists account for a little less than 1/3 of all fatalities. Young men are 8x more likely to die in a car accident than their female counterparts. This "user-hostile" mode of thinking about driving serves no one.

beambot 5y ago

I suspect that a challenging driver assessment between AVs and humans would be the most rapid driver of AVs possible... Just remember: Most drivers think they're above average. Curiously, autonomous car technology also makes it feasible for us to rate drivers -- it's like having a professional watching over your shoulder all the time. IIRC, Tesla used something similar to determine rollout of new FSD capabilities.

nullc 5y ago

The reason to hold humans and self driving cars to a different standard is that humans have already "class qualified" under ... well ... every condition that we drive in... with on the order of a billion driving years of total experience to the limitations of human-driving and far in excess of 10 million years of additional driving added per year along with a considerable amount of study across many decades. Though there are systematic flaws in human drivers they're reasonably well known given the level of experience we have, there are many mitigations in place in both our cars and road designs, and many of these flaws are intuitively modeled by the other humans on the road as well as pedestrians allowing for real-time compensation. The driverless cars, on the other hand-- we have radically less experience with as a class and there are many reasons to believe that vendor to vendor (e.g. due to sensor modalities) or even version to version will behave less self-consistently than human drivers do as a class. We know of some systematic faults in their operation which cause them to behave in ways that are highly surprising to humans, and we should expect to find many more as we gain experience. To the extent that the properties of self driving cars have been formally studied at all it's been primarily by their creators which are obviously self interested. Some of the practices used in early supervised self-driving also actively undermined understanding their safety (e.g. executing a hand over to the human once the car was already in an nonrecoverable unsafe state and then attributing the accident to human driving since at the time of the collision the human was back in charge, but maybe only by a hundred milliseconds; or companies committing perjury to use the DMCA to force down unfavorable videos created by others) This isn't an argument that we might not benefit from more rigorous driver testing or even from testing ideas that originated from self-driving car testing... but parity between driverless car tests and human testing shouldn't be a goal, particularly not when so many orders of magnitude separate our human driving experience from driverless experience.

akira2501 5y ago

There's not a lot of evidence to suggest this would have a major impact. The majority of fatal crashes involve drugs and alcohol and not people with "untested skills in challenging situations." Pedestrians and motorcyclists account for a little less than 1/3 of all fatalities. Young men are 8x more likely to die in a car accident than their female counterparts. This "user-hostile" mode of thinking about driving serves no one.