First Driverless Car Death

Forums › Transport (cars, bikes and boats) › First Driverless Car Death

1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12

28622 posts

Uber Geek

Trusted

Lifetime subscriber

#1983178 26-Mar-2018 09:22

FireEngine:

tdgeek:

It need to have enough sensors, thats not hard. It also needs to have a failsafe. Two sensors everywhere, each controlled by a seperate CPU and seperate power supply. If both fail the cars own ECU will cut power in the case of a autonomous vehicle, if it was a vehicle where it required a human in the drivers seat, it will cut power until the steering wheel etc was manually controlled

There are so many accidents caused by human frailty

Two doesn't do it, you need three and a majority voting system so you can ascertain which of the three is faulty - to do it properly this gets complicated and expensive really fast.

Human frailty affects system design and software coding too. Thats the point. If one individual driver is inattentive then there is usually one accident as a result before corrective action is taken. If a software designer is inattentive on driverless cars then you could have many accidents before it is traced as a root cause and fixed...

Four is overkill. It will be better but it wont exist as its too costly, hence lets cancel self drive permenantly

There is a vast difference between me making a silly human mistake and a coding shortfall. There is not a huge amount of scenarios to test. Its gets coded, its gets tested. When a sensor sees this, it does that. If you brainstormed all the kinds of accidents I could face in real life there is not that many. Something comes at me from ahead at a 90 degree swath (45 each side peripheral). Or comes at me from the side, or the rear. I get a blowout, engine seize, suspension failure and so on. There is not a lot that will face the car, and thus that the software needs to manage. Detect and avoid, following the best means to manage the physics that exist, momentum, traction, what the car can achieve

Or if its seen as too hard to get right, forget it entirely. Let humans make these decisions

Mighty Ape

Shop Mighty Ape for electronics, games, computers books and more (affiliate link).

FireEngine

1217 posts

Uber Geek

Trusted

#1983184 26-Mar-2018 09:31

kryptonjohn:

Yep - and to be honest who would want driverless cars that turn into blind killing machines if something as minor as a headlight bulb blew.

Oh good grief. "Blind killing machines"? Even my old VW can tell me when a bulb's out. You really think a production approved AI based car would carry on with vital equipment malfunctioning?

Thats exactly my point, driverless cars would not be relying on headlights for obstruction detection.

However this particular car <did> carry on regardless of the pedestrian in its path and <is> approved for use on the roads so there is obviously something wrong somewhere...

Regards FireEngine

kryptonjohn

2523 posts

Uber Geek

Lifetime subscriber

#1983185 26-Mar-2018 09:33

It's approved for experimental/trial use, subject to having a human supervisor able to take over. Nothing like a finished product. Not even close.

tdgeek

28622 posts

Uber Geek

Trusted

Lifetime subscriber

#1983186 26-Mar-2018 09:36

FireEngine:

who would want driverless cars that turn into blind killing machines if something as minor as a headlight bulb blew.

I think that sums this topic up. The pro brigade want CPU driving instead of humans as humans are frail drivers. The anti brigade sees any death or injury by a self drive as proof that its a bad idea.

IMO self drive at the moment is well short of testing, this Uber issue should have been detected on a private test road, not in the wild, but it seems authorities didn't see it that way. Maybe they were told it has been tested exhaustively, (wrong) and the human on board will be all eyes (wrong) Whethet this Uber issue was avoidable or not doesnt matter, it has exposed where testing is currently at

FireEngine

1217 posts

Uber Geek

Trusted

#1983187 26-Mar-2018 09:36

tdgeek:

There is a vast difference between me making a silly human mistake and a coding shortfall. There is not a huge amount of scenarios to test. Its gets coded, its gets tested. When a sensor sees this, it does that. If you brainstormed all the kinds of accidents I could face in real life there is not that many. Something comes at me from ahead at a 90 degree swath (45 each side peripheral). Or comes at me from the side, or the rear. I get a blowout, engine seize, suspension failure and so on. There is not a lot that will face the car, and thus that the software needs to manage. Detect and avoid, following the best means to manage the physics that exist, momentum, traction, what the car can achieve

Or if its seen as too hard to get right, forget it entirely. Let humans make these decisions

There is physics (momentum mainly), involved, the threat analysis has to be predictive if it is to work, ie what <may> become an obstruction...and the scenarios for that are vast in degree if not in category. This was also a single threat on an unobstructed road with no other traffic...hence the concern that the car failed in that scenario. I'm not talking a "silly" mistake but it is easy to envisage real-world scenarios that the developers didn't cater far...that is largely the nature of software defects per se.

Regards FireEngine

tdgeek

28622 posts

Uber Geek

Trusted

Lifetime subscriber

#1983196 26-Mar-2018 09:48

FireEngine:

I'm not talking a "silly" mistake but it is easy to envisage real-world scenarios that the developers didn't cater far...that is largely the nature of software defects per se.

Ok, if I sat down and noted as many situations that my car can face in the wild, I pick the number 100. I even pick that I will struggle to get to that number. Thats a very low number to code for. At the lowest level its item A is approaching me and will hit me. Item A may be stationary or moving. There are many other different scenarios but not that many. So we list these 100 . Expert drivers say what should be done to mitigate the accident, taking into account what the car, the tyres the conditions allow for. Its not really rocket science, unless you are a disbeliever

FireEngine

1217 posts

Uber Geek

Trusted

#1983211 26-Mar-2018 10:03

tdgeek:

FireEngine:

I'm not talking a "silly" mistake but it is easy to envisage real-world scenarios that the developers didn't cater far...that is largely the nature of software defects per se.

Ok, if I sat down and noted as many situations that my car can face in the wild, I pick the number 100. I even pick that I will struggle to get to that number. Thats a very low number to code for. At the lowest level its item A is approaching me and will hit me. Item A may be stationary or moving. There are many other different scenarios but not that many. So we list these 100 . Expert drivers say what should be done to mitigate the accident, taking into account what the car, the tyres the conditions allow for. Its not really rocket science, unless you are a disbeliever

Detect item A

Determine what item A is, its currently movement path if any and assess its possible and likely movement paths and speeds...

Detect item B

Determine what item B is, its currently movement path if any and assess its possible and likely movement paths and speeds...

Then determine priority of A over B if required, determine if A and B will interact, does this create an increased likelihood of A or B changing speed and or path to result in correction needed...

No its not rocket science, it is however complicated. Dealing with 1 scenario at a time is straightforward but that is not driving. Driving is potentially tracking the behaviour and possible/likely future behaviour of up to 20 or so vehicles in traffic, at speed. And we did just see a car using the current state of the art technology mow down a pedestrian without hesitation despite on the face of it, it not being a very unlikely scenario...so its clearly not <easy> to code for.

Regards FireEngine

kryptonjohn

2523 posts

Uber Geek

Lifetime subscriber

#1983216 26-Mar-2018 10:12

Of course it's complicated. Ongoing leading technology projects sucking up billions of dollars tend to be.

However nothing in what you raised or what I've seen looks to me to be insurmountable.

Google are in for over US$1B. Toyota are talking about putting $2.8B into it. Uber $1B. Detroit $?B.

This ain't about to go away because it's too hard.

And as I said before, it doesn't be perfect. It just has to kill and maim less often than humans, who do stuff like drive on the wrong side of the road, swerve to avoid possums, drive drunk, drive tired, drive too fast etc.

FireEngine

1217 posts

Uber Geek

Trusted

#1983226 26-Mar-2018 10:28

kryptonjohn:

Of course it's complicated. Ongoing leading technology projects sucking up billions of dollars tend to be.

However nothing in what you raised or what I've seen looks to me to be insurmountable.

Google are in for over US$1B. Toyota are talking about putting $2.8B into it. Uber $1B. Detroit $?B.

This ain't about to go away because it's too hard.

And as I said before, it doesn't be perfect. It just has to kill and maim less often than humans, who do stuff like drive on the wrong side of the road, swerve to avoid possums, drive drunk, drive tired, drive too fast etc.

I'd kinda be more optimistic if it was $4B+ going into a single system, rather than competing systems, as it is several orders of magnitude beyond anything in mainstream cars right now but thats by the by.

This technology will only be taken up on trust, the problem is the "if it does less harm than humans" argument can only be assessed retrospectively. I guess in time the statistics for each make/model will be presented as they are now...?

I go out onto the roads trusting that each and every driver is held accountable for their actions, I'm not sure yet that we are clear who will be held accountable for the actions of a driverless car, that is the opportunity in Arizona right now after this tragedy - to clarify the responsibilities. We'll see if it is taken up or kicked down the road...

Regards FireEngine

kryptonjohn

2523 posts

Uber Geek

Lifetime subscriber

#1983229 26-Mar-2018 10:35

The USA vehicle industry is so regulated and litigious - it will be a very hard road to get this technology approved for general use as well as insurable. Just ask the aviation industry about certification - this will be just has difficult if not more so. I expect they will have to test these things for millions of hours - thousands of cars for thousands of hours, before they get that far. So they may not have as good a perspective on safety as the truly retrospective view, but they will be pretty good idea.

Rikkitic

Awrrr
17246 posts

Uber Geek

Lifetime subscriber

#1983313 26-Mar-2018 11:25

The problem with code is it's not intuitive. It is all based on if this do that. The precision can be increased by breaking actions down further and further, but ultimately it is still if this do that. I wonder if any conceivable AI would have made the decision to put that plane down in the Hudson river.

Plesse igmore amd axxept applogies in adbance fir anu typos

tdgeek

28622 posts

Uber Geek

Trusted

Lifetime subscriber

#1983317 26-Mar-2018 11:34

Rikkitic:

The problem with code is it's not intuitive. It is all based on if this do that. The precision can be increased by breaking actions down further and further, but ultimately it is still if this do that. I wonder if any conceivable AI would have made the decision to put that plane down in the Hudson river.

Self drive isnt AI. But re Sully:

1. Have to land

2. Runway too far based on distance, loss of altitude caused by banking, glide path at given airspeed. Land or water? Its NYC, so water

3, Autopilot lands on water using a "water" setting, i.e. slow as possible, wings level, set it to stall at 6 feet to give lowest possible landing speed, plus keeping nose up, allows engines to catch the water as gradually as possible, dragging the tail section a bit helps slow it down. If wigs not level, one engine grabs and rips the wing off and breaks fuselage, as per the Atlantic hijacking in I think the 80's

tdgeek

28622 posts

Uber Geek

Trusted

Lifetime subscriber

#1983318 26-Mar-2018 11:39

Plus I dont trust human intuition in a car that is 3 seconds from an accident. Pilots know flight dynamics, they are trained for it. Humans know to brake and steer away from danger, often those are both wrong. Even if they are the right choice, the way to brake and the way to steer isn't always foot down and yank the wheel

FireEngine

1217 posts

Uber Geek

Trusted

#1983361 26-Mar-2018 11:52

tdgeek:

Rikkitic:

The problem with code is it's not intuitive. It is all based on if this do that. The precision can be increased by breaking actions down further and further, but ultimately it is still if this do that. I wonder if any conceivable AI would have made the decision to put that plane down in the Hudson river.

Self drive isnt AI. But re Sully:

1. Have to land

2. Runway too far based on distance, loss of altitude caused by banking, glide path at given airspeed. Land or water? Its NYC, so water

3, Autopilot lands on water using a "water" setting, i.e. slow as possible, wings level, set it to stall at 6 feet to give lowest possible landing speed, plus keeping nose up, allows engines to catch the water as gradually as possible, dragging the tail section a bit helps slow it down. If wigs not level, one engine grabs and rips the wing off and breaks fuselage, as per the Atlantic hijacking in I think the 80's

Sully's book is quite interesting, especially the section referring to Goal Sacrifice:

Essentially 1549 or any other flight would have three Flight Goals in order of importance:

1. Passengers Survive

2. Plane Intact

3. Get to Destination

3 went south with the engines, 2 was sacrificed 30secs later after his eyeball of the remaining energy in the plane vs runway locations, <only> Goal 1 then prevails, large flat area of ground not available, hence the Hudson. Quite a simple decision tree when broken down like that...

Amusingly in the movie the camera scans past a section of the dashboard where fleetingly is shown the Airbus' faults display (presumably with the system failures relating to both engines listed), and at the top in red..."Land ASAP" :-)

Regards FireEngine

Rikkitic

Awrrr
17246 posts

Uber Geek

Lifetime subscriber

#1983368 26-Mar-2018 11:57

I wonder if the autopilot includes an instruction Do NOT land on water.

Plesse igmore amd axxept applogies in adbance fir anu typos

1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12