Hacker News Reader

Comment by conception | Hacker News Reader

Normal drivers, Nuthead bikers, drunk pedestrians and cops can all be reacted to much swifter by a computer using radar than any driver behind the wheel. It's just pride to think you could.

View on HN

I don't think anyone doubts it can react more quickly. The real question is whether it can react more correctly. I think the answer is "soon", for what that's worth, but I also think that until the answer is an unmitigated "yes" it's a valid criticism.

(edited for tone)

Well, whether it acts more correctly becomes a moot point when it acts correct most of the time, then it's about what acts more safely. There are advantages to having a system that's not allowed to be lax when it comes to safety rules.

The key is that a human can anticipate a dangerous situation and take mitigating action before a quick reaction is necessary. AI still has far to go before that sort of nuanced behavior can be replicated. Until then self driving cars will be like teenagers learning to drive. Sure they can get themselves from point A to B but I don't want to be near them while they do it.

Disagree entirely. Anticipation is just a function of memory - of having enough training data that you can glean potential outcomes and their probabilities through the attributes that surround them.

That's used in AI all the time already.

You're drivng down a crowded highway and you see an obstacle. You're likely to get rear ended if you hit your brakes and you can't change lanes. However you recognize it's just an empty card oard box. You drive through it with no issues.

Can you train a computer to do that?

...and it turns out the "empty" cardboard box is half filled with metal pipes, which your human eye had no way of knowing in the split-second before your car collides into it in 65 mph. Your car screeches, loses control, and swerves into the neighboring lane.

But luckily a following self-driving car already detected the presence of an obstacle using radar. It detects your car maintaining speed. Within several milliseconds it evaluates chance of collision as VERY HIGH, starts evasive maneuver, scratching your car at a speed which is still uncomfortably high but not likely to cause a fatal accident.

Several weeks later a team of automobile experts review everything recorded in the accident, improve the algorithm to reduce the chance of accident even further in Edge Case 783645: the preceding human-driving car plowing into a small road obstacle at 65 mph. Firmwares are automatically upgraded during the following months.

to recognize a cardboard box? yes.

The computer could immediately check adjacent lanes, communicate with the car behind, know how much space is on the shoulder, take information from other cars who've driven beside the box already, etc. In a fraction of a second.

AI cars could communicate amongst themselves to forewarn each other. They would instantly know typical driving patterns for any location at particular times and in particular conditions. A human would know that for their usual routes at best.

Instincts are strange. I have been (happily) surprised by myself on multiple occasions by how I've reacted to near-impact situations. Do I think that means cars can't do it? Absolutely not. I do think that it will take a lot of learning to get the same out of them. On the other hand, do I also think that replacing every car on the road with a self-driving one right now would result in fewer accidents and fatalities? Yes. If it were illegal to drive and all vehicles were required to be autonomous, I 100% believe that would drop the danger of motor vehicles significantly. Where it gets weird is the concept of fault. Right now, because most motor vehicle accidents are driver error, humans screwing up almost serves as an insurance policy for manufacturers. If fewer people die, but more of them die as a result of manufacturers' software... Yeah I don't know. The future is going to be weird.

Well, I'd say it's just pride to be so confident in the ability of human engineers to make machines outperform humans consistently (at least in the next few years that vendors want to start selling auto-drivers.) I don't think anyone's doubting whether radar+computer can move faster than a human, but we may doubt whether it will make the "right" reaction and if it might make unwanted decisions in some other cases we don't think of until they happen. And that's just the "AI" level, not stupid things like your electronic throttle's stack overflowing and locking in the open position[1].

[1] http://embeddedgurus.com/state-space/2014/02/are-we-shooting...

You speak as if humans are really good at making driving decisions today. We're really not.

I didn't exactly say that. But I think you and others are underestimating our collective skills. We make stupid decisions, get tired, etc., but we manage to recover, workaround others' mistakes, etc. I think people are underestimating how hard it will be to get "the last 10%" of driving skill, that gap between slapping together cruise control+camera+radar+GPS to be able to handle the best-case scenarios, and being able to drive at least as well as the average human in the average-case scenarios.

Another thought: what if it takes a generation or longer for self-driving cars to do well with cases like slippery rain or ice? People will stop practicing how to drive at all, until suddenly one day they have to drive in the most dangerous conditions possible. Will the net cost/casualties of accidents still be lower than it is today?

I suppose it depends on what your standard is and what you're comparing it with. The Insurance Institute for highway safety says there is about 1 fatality per 100 million vehicle miles travelled in the US.

http://www.iihs.org/iihs/topics/t/general-statistics/fatalit...

That still translates into a lot of annual deaths of course (and doesn't capture the many serious injuries not included in that statistic). But humans actually drive a lot of miles with relatively few serious accidents.

Swifter, yes. Better? I think that depends on the particular circumstance. For example: you look in your rear view mirror and notice an out of control cement truck half a block away heading directly towards you, plowing through vehicles in its path. Would a computer recognize the impending catastrophe and take the initiative, sideswiping a few vehicles in your way in an effort to get out of the situation?

I think it's safe to say that every day driving circumstances are already on par with what a human driver could achieve. It's the extenuating circumstances where a human can make the call of "my life is more important than property damage or even risking others lives" where I have my doubts. Those situations are obviously rare, but still quite important.

If they're extremely rare, then they're not important.

If you save X lives in every day driving circumstances at the cost of Y (Y << X) lives in bizarre rare circumstances, it's an easy, obvious win.

That may be the rational response under certain morals, but not everybody is rational always or has the same morals. Who is legally responsible for those Y deaths? Or, even if there was no death but the car broke a traffic law, how is that addressed? If the Y deaths are disproportionately a distinct class of people from X (say, pedestrians, construction workers), are they going to be eligible for some class action lawsuit? Or will it deepen class divides as poor pedestrians become bitter about being hit by self-driving cars carrying rich people? Are the "drivers" going to suffer guilt, wondering if there's something they could have done to prevent an accident? Are issues like this going to cause the law to require a human "standby driver" ready to take over, canceling many of the hoped-for advantages of self-driving cars? If drivers have legal or monetary liability, are they going to want to hand that over to a machine built as cheaply as possible and designed by overworked startup engineers? (perhaps the first generations at least.) Will the cars keep audit records to prove "who" was driving, etc.?

Maybe some of this is getting off-topic; I'm just musing on how complex the practical and societal factors could get.

>Are issues like this going to cause the law to require a human "standby driver" ready to take over, canceling many of the hoped-for advantages of self-driving cars?

The law can say whatever it wants but there's effectively no such thing as a "standby" (for <30+ second timeframes) for a driving system that's operating autonomously. The driver is watching a movie, reading a book, or sleeping. And if the driver isn't allowed to do any of those things--say by the system shutting down if they take their eye off the road for more than 5 seconds--no, then, the system is largely useless.

In which case people won't pay for it except to the degree that it's essentially an assistive driving safety system.

Mhmmm no. If you have a machine that is manually operated and kills 1/1000 people due to operator errors, and a machine that is fully automatic but kills 1/10000 people due to software errors or "weird edge cases", it's still the second machine that will never be allowed on the market, because even one instance of it killing anyone will cause it to be fully recalled - doesn't matter that statistically it kills less people than a manually operated one. For automatic machines that number is 0, and automatic cars are unlikely to achieve that.

What about in bad weather?

Reply on news.ycombinator.com