3 ms·

This is a hard problem because driving software is so complicated with many special cases. Life and safety code in general is hard. Do you want to mandate they

by imglorp 10y ago

This is a hard problem because driving software is so complicated with many special cases. Life and safety code in general is hard. Do you want to mandate they apply aerospace coding technology? Like the other guy said, this could set them back 20 years. Mandate a traceability culture like ISO9000? How about third party audits? Escrow the code and mandate telemetry for NTSB to examine after an incident? Even if there are a few bugs, they'll still be safer than humans. Google's up to 1.5 million miles driven with no fatalities. The human metric to beat will be 1.2 fatalities per 100M miles.

kalleboo 10y ago
> Google's up to 1.5 million miles driven with no fatalities. The human metric to beat will be 1.2 fatalities per 100M miles I thought Google's speed limit was 25 mph? What's the human metric for fatalities at 25 mph?
- ocdtrekkie 10y ago
  Google's own designed cars are, not the Lexus models. But their stats are junk stats in another way: Test drivers simply intervene super regularly to avoid any sort of autonomous incident.
  - jacquesm 10y ago
    This is a very hard problem to solve. The longer the interval between interventions the bigger the chance that the 'driver' will not be able to intervene in time because they no longer have situational awareness.
    - ocdtrekkie 10y ago
      I think the best judge of self-driving aptitude, though, would be "average contiguous self-driving miles driven". The total miles driven figure doesn't mean anything if a driver has to intervene every time you approach an intersection or something. Every time the test driver intervenes, reset the clock.
paganel 10y ago
> Google's up to 1.5 million miles driven with no fatalities. Has the Google car driven on a Bulgarian road late at night, rain pouring down, with no road-markups, with lorries coming your way and blinding you every couple of minutes? I did that yesterday, and after those moments when I was trying to guess "what is the actual road?" I was thinking that there's no way in hell a computer program would be able to manage that unless full-AI becomes a thing.
- imglorp 10y ago
  That's where crowdsourced data comes in. But of course it doesn't help if the road moves or there's construction or a rockslide or something.
- ams6110 10y ago
  I did that last night in the USA. Well it wasn't raining -- but an unpaved, washboarded road with no lane markings of any kind, blind hill crests and turns, and just barely wide enough for two cars to pass.
- icebraining 10y ago
  While it's certainly hard - and remember, people don't manage that either, at least not as safely as we will require self-driving cars to operate - I don't see why it'd need "full-AI". Recognizing the road is not a matter of general intelligence, it's a matter of having very good but very specific recognition abilities. Besides, remember that a self-driving car won't have to operate with just a couple of eyes stuck inside a cabin. Lorries won't blind radar.
- jamez1 10y ago
  Do you have any idea how self driving cars work? They currently have sensors that are much different from your eyes behind a windscreen, they don't get blinded by light and rain doesn't bother them.

by imglorp 10y ago

kalleboo 10y ago

> Google's up to 1.5 million miles driven with no fatalities. The human metric to beat will be 1.2 fatalities per 100M miles I thought Google's speed limit was 25 mph? What's the human metric for fatalities at 25 mph?

ocdtrekkie 10y ago
Google's own designed cars are, not the Lexus models. But their stats are junk stats in another way: Test drivers simply intervene super regularly to avoid any sort of autonomous incident.
- jacquesm 10y ago
  This is a very hard problem to solve. The longer the interval between interventions the bigger the chance that the 'driver' will not be able to intervene in time because they no longer have situational awareness.
  - ocdtrekkie 10y ago
    I think the best judge of self-driving aptitude, though, would be "average contiguous self-driving miles driven". The total miles driven figure doesn't mean anything if a driver has to intervene every time you approach an intersection or something. Every time the test driver intervenes, reset the clock.

paganel 10y ago

> Google's up to 1.5 million miles driven with no fatalities. Has the Google car driven on a Bulgarian road late at night, rain pouring down, with no road-markups, with lorries coming your way and blinding you every couple of minutes? I did that yesterday, and after those moments when I was trying to guess "what is the actual road?" I was thinking that there's no way in hell a computer program would be able to manage that unless full-AI becomes a thing.

imglorp 10y ago
That's where crowdsourced data comes in. But of course it doesn't help if the road moves or there's construction or a rockslide or something.
ams6110 10y ago
I did that last night in the USA. Well it wasn't raining -- but an unpaved, washboarded road with no lane markings of any kind, blind hill crests and turns, and just barely wide enough for two cars to pass.
icebraining 10y ago
While it's certainly hard - and remember, people don't manage that either, at least not as safely as we will require self-driving cars to operate - I don't see why it'd need "full-AI". Recognizing the road is not a matter of general intelligence, it's a matter of having very good but very specific recognition abilities. Besides, remember that a self-driving car won't have to operate with just a couple of eyes stuck inside a cabin. Lorries won't blind radar.
jamez1 10y ago
Do you have any idea how self driving cars work? They currently have sensors that are much different from your eyes behind a windscreen, they don't get blinded by light and rain doesn't bother them.