Page 5 of 6

Re: Safety systems are there to be gamed...

Posted: Mon Jan 29, 2018 2:24 pm
by Horse
ancient wrote:
TR4ffic wrote:It shouldn't be too difficult to think up a scenario for an impending collision where the autonomous car you are travelling in will sacrifice your life (or at least seriously injure you) rather than taking evasive action that would/could maim or kill others.

Can you come up with one?


Approaching a tunnel. Several pedestrians appear in the tunnel, out of sight of the vehicle until it is within stopping distance. The vehicle is aware of multiple pedestrians and one vehicle occupant, so attempts:
- Emergency braking, to slow (stopping not possible)
- Avoidance manoeuvre, to miss unprotected pedestrians
AV impacts side of tunnel portal
AV occupant may have injury severity reduced by secondary safety (crumple zones, seatbelts, airbags, etc.


OK, that's slightly different to than taking evasive action that would/could maim or kill others, but similar principle?

Re: Safety systems are there to be gamed...

Posted: Mon Jan 29, 2018 2:57 pm
by TR4ffic
OK. It's all "what if's" but it's the principle I wanted to debate/get across

1. Your AV and you are in a built up area. You've just left a junction going up hill - in the opposite direction to your right a stationary queue of vehicles waiting at the junction you've just left and to your left is a pavement full of pedestrians. An HGV coming down the hill, for whatever reason, is going too fast, fails to stop at the end of the queue and crosses onto your side of the road. There's nowhere to go so your AV simply stops leaving you to your fate and whatever protection your AV can give you... I don't know, but I expect your average human driver would take avoiding action to the left onto the pavement regardless of the pedestrians there.

2. Your AV and you are on a multi-lane carriageway with no central reservation (The A38M into/out of Birmingham from Spaghetti Junction springs to mind) - You're in the R/H lane for the traffic travelling in your direction, in a steady stream of traffic, cars and a couple of motorcycles in the lane to your left. A vehicle coming the opposite way looses control and is on a collision course for you. Taking avoiding action to the left would/could mean colliding with other vehicles/motorcycles- taking avoiding action to the right would mean coming into conflict with other vehicles travelling in the opposite direction. Does your AV just elect to take the collision? On the other hand, your AV could possibly just brake or even accelerate and the out of control vehicle could pass harmlessly (to you) either just in front or just behind you, and everyone else can take their chances.

The point is that a human driver will take evasive action to avoid the immediate collision with little or no regard for the consequences - It's human nature ...but it won't be 'AV nature'...

Re: Safety systems are there to be gamed...

Posted: Mon Jan 29, 2018 3:57 pm
by ancient
Horse wrote:
ancient wrote:
TR4ffic wrote:It shouldn't be too difficult to think up a scenario for an impending collision where the autonomous car you are travelling in will sacrifice your life (or at least seriously injure you) rather than taking evasive action that would/could maim or kill others.

Can you come up with one?


Approaching a tunnel. Several pedestrians appear in the tunnel, out of sight of the vehicle until it is within stopping distance. The vehicle is aware of multiple pedestrians and one vehicle occupant, so attempts:
- Emergency braking, to slow (stopping not possible)
- Avoidance manoeuvre, to miss unprotected pedestrians
AV impacts side of tunnel portal
AV occupant may have injury severity reduced by secondary safety (crumple zones, seatbelts, airbags, etc.


OK, that's slightly different to than taking evasive action that would/could maim or kill others, but similar principle?

Surely the AV will not be travelling at a speed where it cannot stop (safely, on it's own side of the road) within the distance that it can 'see' (detect) to be clear. So it would have slowed before entering the tunnel and the pedestrians could not appear >within< stopping distance, but would appear (at most) at the limit of safe stopping distance.

Re: Safety systems are there to be gamed...

Posted: Mon Jan 29, 2018 4:44 pm
by Horse
ancient wrote:
Horse wrote:
ancient wrote:
TR4ffic wrote:It shouldn't be too difficult to think up a scenario for an impending collision where the autonomous car you are travelling in will sacrifice your life (or at least seriously injure you) rather than taking evasive action that would/could maim or kill others.

Can you come up with one?
OK, that's slightly different to than taking evasive action that would/could maim or kill others, but similar principle?

Surely the AV will not be travelling at a speed where it cannot stop (safely, on it's own side of the road) within the distance that it can 'see' (detect) to be clear. So it would have slowed before entering the tunnel and the pedestrians could not appear >within< stopping distance, but would appear (at most) at the limit of safe stopping distance.


Always and never. Two words you should always remember to never use.

Re: Safety systems are there to be gamed...

Posted: Tue Jan 30, 2018 9:55 am
by dvenman
If human drivers can't get it right all the time, how do we expect autonomous vehicles to do the same ?

Re: Safety systems are there to be gamed...

Posted: Tue Jan 30, 2018 10:44 am
by Horse
dvenman wrote:If human drivers can't get it right all the time, how do we expect autonomous vehicles to do the same ?


That's AVs running on software written by humans :)

Re: Safety systems are there to be gamed...

Posted: Tue Jan 30, 2018 11:16 am
by Horse
Cat :)

I was mulling over this while being chauffeured on a night out. Our driver was whizzing happily along a residential road, about 10:30pm, doing 20-25mph. There was a cat sat under a parked car (nope, doesn't scan like Dr Suess), watching us approach. Given the possibility of it running across in front, I'd have slowed. Our driver didn't.

Drop-off done, we returned up the road. Cat just visible under the car, and luckily it stayed there.

Did my driver see it? Well, it was clearly visible and - AFAIK - he's not blind. But he didn't acknowledge its presence or alter his driving to allow for random cat avoidance.

Perhaps this is a good simile for the realistic, immediate, future of AV technology, with the DVSA hazard Perception Test giving a parallel of how the AVs will drive? In tat test, the marking isn't based on 'hazard' as I was taught it in old 'blue book' days, of 'actual or potential danger'. Instead, HPT marks candidates on reaction to a 'developing' hazard, one that needs action by the driver.

So the AV's lidar might have 'seen' the cat, identifying it as a variant from it's mapping (as, for that matter, would the parked car have been a variant). But it *might* not have reacted unless the cat moved. However, where it *might* be a step above the human HPT graduate driver is that it *might* have slowed given the possibility of anything emerging from beyond the parked vehicle.

Under, over, through. We might look, see, identify, predict, react - before something happens. Is that the 'eventual aim' for AV designers?

Actually, yes. Some I heard give presentations recently said that their aim is 'advanced driving' - they are aware of it and would like their AVs to drive like that.

Re: Safety systems are there to be gamed...

Posted: Tue Jan 30, 2018 11:23 am
by Horse
jont- wrote: :lol: at 5mph. Would almost be quicker just to walk.


The pods have a maximum capable speed of 24kmh (15mph) but will be limited electronically depending upon the environment they are travelling in (for example, moving more slowly in congested areas). In general, they will move in pedestrianised areas at a brisk walking pace.

https://ts.catapult.org.uk/innovation-c ... oject-faq/

Re: Safety systems are there to be gamed...

Posted: Thu Feb 01, 2018 3:20 pm
by crr003
Horse wrote:
dvenman wrote:If human drivers can't get it right all the time, how do we expect autonomous vehicles to do the same ?


That's AVs running on software written by humans :)

Isn't the idea to have AI "learn" as it matures, so your example of a pussy might get a speed reduction/position change, but after 25 pussies who have just sat there, the AI computes the risk probability/impact as low and just cracks on? Besides, not one of the eight, so not a problem to hit (and I like pussies).

Image

Anyway, has no one watched the film Ex Machina? That's the AI future..................... :ugeek: :shock:

Re: Safety systems are there to be gamed...

Posted: Thu Feb 01, 2018 4:02 pm
by Horse
crr003 wrote: Isn't the idea to have AI "learn" as it matures, so your example of a pussy might get a speed reduction/position change . . .


TBH, I don't know. Alternatively, it might identify it has a heat source (if it can differentiate from a hot exhaust) and so still consider it higher risk? I'm guessing :)

crr003 wrote: Besides, not one of the eight, so not a problem to hit


Schroedinger used cats because they're a cost-efficient lab animal - 9 lives :)