Driving is a social activity. You’re interacting with other humans, with only their behavior (and maybe the occasional gesture) from which to infer intent.
From that perspective, think of what a computer has to “know” in order to get along? It’s way more complex than just following the road and not hitting things.
Maybe the best chance for a fully autonomous driving experience is some city that takes the next step beyond congestion pricing and reserves their dense core for autonomous vehicles. If computers only have to deal with other computers, and pedestrians, that seems a much more tractable problem.
they'll be on Motorways first, there's just so many less edge cases.
Not from what I've seen. Heavy construction, sudden gridlock with huge lane speed differentials, multi-vehicle collisions, road debris, heavy trucks, severe inclement weather, high speeds, gore points; these things combine to make for an extremely challenging and dangerous driving environment.
An autonomous car driving through a residential neighbourhood is moving slowly and can stop if a child chases a basketball onto the road. On the freeway, when you're going over 100km/h, the car simply can't stop if one from the next lane spins out in front of you.
> On the freeway, when you're going over 100km/h, the car simply can't stop if one from the next lane spins out in front of you.
Neither can you.
In fact, an autonomous car is much more likely to be able to react fast and well to save lives and property in that situation than a human.
Self-driving cars don't need to be absolutely 100% perfect and accident-free to be worthwhile. They just need to be better than us, and frankly, we're pretty lousy at operating tons of metal flying along the road at over 100km/h. Hell, we're not even that great at it when going 50 km/h.
Humans have the advantage of understanding context enough to (try to) avoid being in those situations though. You may not be able to avoid a car suddenly spinning out right in front of the you, but you may have realized 7 minutes ago the driver kept drifting outside his lane and backed off to give him a bigger gap; getting autonomous cars up to that level would require strong AI tech
Not at all. That level of pattern recognition is much easier than the actual hard problems involved in getting driverless cars working.
The hard problems, right now, are things like recognizing where the road is when it (and everything beside it) is under an inch and a half of snow, or where you're supposed to drive when there's construction and the lanes are shifted (assuming no new standard means are developed to indicate this), or how to recognize, and what to do, when the road you want to take is under a foot of rushing water—or washed out altogether.
This is the thing that continues to baffle me. There are genuine, hard problems between where we are now and a nation of completely autonomous cars. But every time there's a discussion about it, even on a site like Hacker News where people should have the background to recognize the difference, most of the problems people bring up are the kind that self-driving cars either are now or can fairly easily become good at dealing with.
People say this, but I suspect that "epsilon better" is not enough. Personally, I think that to give up my (fallble) agency, I would require an order of magnitude better, perhaps two.
From that perspective, think of what a computer has to “know” in order to get along? It’s way more complex than just following the road and not hitting things.
Maybe the best chance for a fully autonomous driving experience is some city that takes the next step beyond congestion pricing and reserves their dense core for autonomous vehicles. If computers only have to deal with other computers, and pedestrians, that seems a much more tractable problem.