Hacker News Reader

Comment by elmin | Hacker News Reader

This seems like marketing-speak to me. Or perhaps they haven't gotten far enough in their development for it to be really hard yet. Here are some examples of reasons this is harder than you'd think:

- At a complex intersection, how does the computer decide which traffic light is yours?

- You see an obstruction in the road. How do you decide if it's something like a plastic bag you can drive through, or something you need to swerve around?

A very experienced team has been working on this at Google for almost a decade. I'm not sure how Elon's team would have skipped so many steps.

View on HN

Indeed,

And even the google car ignores entirely the traffic lights that aren't on the highly detailed map guiding the car [edit: that disappears when it loses connectivity. But connectivity is a "solved problem" too, right?].

But the simple parts of the problem are solvable. The very simplest parts of the problem could probably be solved with analog relays.

One theory: get enough hype and support for autonomous cars and cities will be forced to install wireless guides for autonomous cars. Then the only problem is people stepping in front of them. So you ignore that problem and away you go.

I mean, regular automobiles, when they were first introduced, caused massive casualties. Only in the eighties did we begin to get this under control with air bags, sane speed limits and so-forth.

Cars have been just rolling death for a large portion of the history of people using them. Well, autonomous cars can just leverage that fine history.

We haven't been improving on drivers, though. Twenty years ago, humans as a group were as bad as they are today and as they will be in another twenty years. Or thirty or fifty. You can add seatbelts and rear parking cameras, and you can legislate against texting while driving, but you can't legislate away tired or distracted drivers, or just plain unskilled drivers. As a group, the only way we improve is through demographic shift. But autonomous cars get better, as a whole, and consistently, and rapidly.

Maybe autonomous cars will be at the 50th percentile in five years, but they're only as good as the bottom 10% of drivers, well, we'd still be better off with those 10% of drivers switching--which would lead to infrastructure to improve it. (I think more importantly, social understanding/recognition of them. The hard problems presented in Dr. Leonard's video posted on this thread aren't peds stepping in front of them, but cops gesturing and unprotected left turns across traffic where you need to make eye contact.)

Finding the bottom 10% sounds hard, because, yes, a lot of bad drivers think they're good drivers, but I think the majority of us know we suck. For many, their insurance companies know too.

For the traffic light problem, I'd start off by programming the car with common scenarios, and when you hit one it doesn't understand, throw up a prompt that tells the driver to tap the accelerator when it's ready to go. You can then improve from there. "Driverless" doesn't have to be all or nothing. We can ease into it.

For the plastic bag, I find it odd how often people bring up scenarios like this that will supposedly stump a driverless car, yet humans are terrible at as well. "What if a massive truck comes careening down the road straight at you, and your only avenue of escape is through a playground full of children? What does the car do?" Well, somebody dies, just like now. My experience is that humans are terrible at deciding whether some crap in the road is safe to run over, so computers can hardly do worse. And I think they could easily do better. A front-facing radar (already standard equipment on many cars) will tell you whether that lump the camera sees is actually substantial or just filled with air.

"What if a massive truck comes careening down the road straight at you, and your only avenue of escape is through a playground full of children? What does the car do?"

We are going to let a computer make that decision? I don't think so.

Why not? A computer can make a much better decision than you do (doing a proper evaluation of options and probabilities) and then execute it much better than a human ever could.

I mean, not pushing for autonomous driving just feels... weird.

The thing is, not only will we let a computer make the decision, but I'm sure we already have.

Automatic collision avoidance is a real-world feature. There are real cars out there you can buy which you can drive straight at a brick wall with the accelerator to the floor and it will not let you hit. I'm sure that avoiding collisions with pedestrians (or big trucks) is part of the deal. If the correct answer isn't "brake" then the computer will get the wrong answer, and almost certainly has already somewhere in the world.

People discuss autonomous driving as basically flipping a switch. We have manually driven cars, and then one day we'll wake up and there will be an autonomous car on the road. But it won't be like that. Autonomy will be a gradual process. Car makers are steadily bundling in more and more automatic driving features. Eventually they'll fill in all the gaps, one by one.

Reply on news.ycombinator.com