Register a SA Forums Account here!
JOINING THE SA FORUMS WILL REMOVE THIS BIG AD, THE ANNOYING UNDERLINED ADS, AND STUPID INTERSTITIAL ADS!!!

You can: log in, read the tech support FAQ, or request your lost password. This dumb message (and those ads) will appear on every screen until you register! Get rid of this crap by registering your own SA Forums Account and joining roughly 150,000 Goons, for the one-time price of $9.95! We charge money because it costs us money per month for bills, and since we don't believe in showing ads to our users, we try to make the money back through forum registrations.
 
  • Post
  • Reply
Improbable Lobster
Jan 6, 2012

"From each according to his ability" said Ares. It sounded like a quotation.
Buglord

Megabound posted:

No, but I would trust a tuned, trialled and rigorously tested program.

Selecting a emergency landing site would not be the hardest problem in the world. One possible solution would be giving the plane a list of green zones where it can land along with the map data it already has. It already has access to its own location via gps.

Or you could give it a downward facing camera. It knows its elevation and your camera is of known specifications, combined with a edge detection algorithm you could identify large contiguous areas. Give the plane a minimum area to shoot for, and it could take care of the rest.

Algorithms aren't a magic spell that makes computers work gooder

Adbot
ADBOT LOVES YOU

Megabound
Oct 20, 2012

Improbable Lobster posted:

Algorithms aren't a magic spell that makes computers work gooder

No, they're combined in specifically designed solutions for executing tasks reliably.

Lime Tonics
Nov 7, 2015

by FactsAreUseless
More russian haircuts,

http://i.imgur.com/XEDdlRR.gifv

hanales
Nov 3, 2013

I feel like split ends would be a problem.

Maxwells Demon
Jan 15, 2007



A real hatchet job.

oohhboy
Jun 8, 2013

by Jeffrey of YOSPOS

Lurking Haro posted:

If a company decides to use Windows for their autonomous whatever, it's clearly their fault if anything goes wrong.

And thank you for going down the trolley rail again.
I'm pretty sure that any plane accident killing hundred on the ground would also kill the passengers.

If not Windows, OSX? Linux? It doesn't matter. The core of the question is do you trust that the software won't bug out on you. You are not talking about the Space Shuttle programmed to fly a nice and tight profile with a machine of known performance bug tested to the point it is a truly bug free program. Unlike a word document if the computer on the plane crashes auto save is not going to help you.

Good going referencing the trolley problem and not understanding it by throwing up your hands and declaring everyone dead. Yay third way!

Megabound posted:

No, but I would trust a tuned, trialled and rigorously tested program.

Selecting a emergency landing site would not be the hardest problem in the world. One possible solution would be giving the plane a list of green zones where it can land along with the map data it already has. It already has access to its own location via gps.

Or you could give it a downward facing camera. It knows its elevation and your camera is of known specifications, combined with a edge detection algorithm you could identify large contiguous areas. Give the plane a minimum area to shoot for, and it could take care of the rest.

You know you are not going to get this program.

The US military has an unlimited budget to design software for 1 plane and they routinely gently caress that up. What are the chances that someone would come up with a "Plane OS" that can cover so many variants of aircraft with so many different internal equipment and histories. Hence why I referenced Windows as it has to deal with the same issue and MS also has near unlimited resources they can tap.

Even in a closed system like iOS where the variants are known and measured on 2 hands craps out too much for such a critical application.

Just finding large contiguous area would get you killed. You would need to identify the material the ground is made of. Is it ploughed? Crops? Ice? Camera knocked out, iced over or partial electrical failure rendering a bunch of sensors blind? What about people? What happens if there are no green zones like airports in middle of cities? What if there is no emergency at all but the plane thinks so?

Megabound posted:

No, they're combined in specifically designed solutions for executing tasks reliably.

Another great way to get killed. Here is a non-harmful example of a bunch of seemingly harmless algorithms interacting to cause unexpected results. Each item you add drives up the complexity massively which in turn increases points of failure.

https://www.youtube.com/watch?v=6yWf6BHqiWM

I would trust a automated car in time as it has a easy clearly defined acceptable fail state, but I wouldn't do that same for a plane as there is far too many failure modes that I know a computer can't handle or be programmed to assuming the failure isn't the computer itself.

Lurking Haro posted:

Way to go, equalling drive-assist systems to autonomous ones.
It should be mandatory that an autonomous car should refuse to work or at least drastically limit its performance if any sensor is giving inplausible data.
It will be 100% the owners fault if anything happens due to neglected maintenance.
Don't tell anyone but planes fly with defects and broken equipment all the time. They get noted down tech log/defect sheet and are sorted in the next maintenance cycle.

Gromit
Aug 15, 2000

I am an oppressed White Male, Asian women wont serve me! Save me Campbell Newman!!!!!!!

oohhboy posted:

If not Windows, OSX? Linux? It doesn't matter. The core of the question is do you trust that the software won't bug out on you. You are not talking about the Space Shuttle programmed to fly a nice and tight profile with a machine of known performance bug tested to the point it is a truly bug free program. Unlike a word document if the computer on the plane crashes auto save is not going to help you.

I don't have a clue so this is an honest question: what is the current state of software control in modern cars? Do they have control over the engine and braking system such that they could accelerate you down the street and plough into oncoming traffic? If so, loads of us are trusting software with our lives.

When it comes to autonomous vehicles, I'll be happy if they perform as good as humans. Performing better would just be icing on the cake. Over 30,000 people die on the roads in the US each year, the bulk of which is a result of driver behaviour. I wonder how many deaths the computers will be allowed before they are considered a failure?

MrYenko
Jun 18, 2012

#2 isn't ALWAYS bad...

Gromit posted:

I don't have a clue so this is an honest question: what is the current state of software control in modern cars? Do they have control over the engine and braking system such that they could accelerate you down the street and plough into oncoming traffic? If so, loads of us are trusting software with our lives.

Throttle-by-wire has been a thing since AT LEAST 2003 or so. The issue is that the OEs are allowing the LTE/4G connection access to the ECM/BCM, which is loving idiotic. I think it was Chrysler that had an incident where people had hacked into the BCM of some Jeep models, and could apply brakes remotely.

GotLag
Jul 17, 2005

食べちゃダメだよ
What if the autonomous truck doesn't realise it's over-height and goes down a tunnel it's too big for?

What if the driving software locks up and it drives into a motorcyclist?

What if the on-board computer has insufficient resources and fails to calculate its course in time?

What if the software drives too fast to negotiate a corner?

What if the computer is on amphetamines?

GotLag fucked around with this message at 01:06 on Apr 23, 2017

jsoh
Mar 24, 2007

O Muhammad, I seek your intercession with my Lord for the return of my eyesight
the dang wiki leak of the cia from like last month confirmed that they are capable of destroying your car remotely if it has network capabilities. never drive in a networked vehicle

Jabor
Jul 16, 2010

#1 Loser at SpaceChem
hackers could turn your car into a BOMB

CAPTAIN CAPSLOCK
Sep 11, 2001



We are finally approaching the point to where I can download a car. :filez:

Tumble
Jun 24, 2003
I'm not thinking of anything!
lol that anybody here thinks they'd be interesting enough for anybody to ever hack

"guys a very small number of people know of vulnerabilities that take a ton of technical knowledge to act on, therefore modernizing cars is bad"

Gumbel2Gumbel
Apr 28, 2010

Tumble posted:

lol that anybody here thinks they'd be interesting enough for anybody to ever hack

"guys a very small number of people know of vulnerabilities that take a ton of technical knowledge to act on, therefore modernizing cars is bad"

Yeah imagine the Internet of Things but for all Toyotas and it makes them crash.

jsoh
Mar 24, 2007

O Muhammad, I seek your intercession with my Lord for the return of my eyesight

Tumble posted:

lol that anybody here thinks they'd be interesting enough for anybody to ever hack

"guys a very small number of people know of vulnerabilities that take a ton of technical knowledge to act on, therefore modernizing cars is bad"

yeah its not like there's any value in someone controlling your brakes to demand money from you right. no one would threaten someone with violence for money

jsoh
Mar 24, 2007

O Muhammad, I seek your intercession with my Lord for the return of my eyesight
part of the point of that wiki leaks thing on the cia was that most or all of those tools they developed are in the wild right now. the things are available

Semisponge
Mar 9, 2006

I FUCKING LOVE BUTTS

CAPTAIN CAPSLOCK posted:

We are finally approaching the point to where I can download a car. :filez:
Farmers have been jailbreaking John Deere equipment because its software is so restrictive they can't perform their own maintenance.

http://boingboing.net/2017/03/22/make-hay-while-the-sun-shines.html

AreWeDrunkYet
Jul 8, 2006

MausoleumExtremist posted:

It will most likely get delayed for decades by transportation lobbyists/UAW/Teamsters etc. Additionally, for the same reason we don't have self-flying commercial jets it'll at most result in having a driver in the seat as a backup.

We have the technology to take off, navigate, and land planes autonomously, but people want some dude in the front seat calmly announcing the current time in Duluth after pressing the LAND button. The same mindset will keep commercial drivers in the seats, especially with the help of the lobbyists.

It takes probably costs on the order of magnitude of what, a couple thousand dollars or so per flight to have a pilot's and copilot's asses in those seats? That seems reasonable as a back-up in case the automated systems fail, plus they handle other miscellaneous responsibilities that would still have to be handled by other employees which is at least a partial offset to the cost.

AreWeDrunkYet fucked around with this message at 01:48 on Apr 23, 2017

Improbable Lobster
Jan 6, 2012

"From each according to his ability" said Ares. It sounded like a quotation.
Buglord

AreWeDrunkYet posted:

It takes probably costs on the order of magnitude of what, a couple thousand dollars or so per flight to have a pilot's and copilot's asses in those seats? That seems reasonable as a back-up in case the automated systems fail, plus they handle other miscellaneous responsibilities that would still have to be handled by other employees which is at least a partial offset to the cost.

Hell, you could even let the pilot and copilot fly it and save thousands that you would have spent installing the computer

Blitter
Mar 16, 2011

Intellectual
AI Enthusiast

oohhboy posted:

If not Windows, OSX? Linux? It doesn't matter. The core of the question is do you trust that the software won't bug out on you. You are not talking about the Space Shuttle programmed to fly a nice and tight profile with a machine of known performance bug tested to the point it is a truly bug free program. Unlike a word document if the computer on the plane crashes auto save is not going to help you.

Anything that is going to run an fully autonomous machine that can kill people will have to be safety certified to an existing standard. An example would be avionics software certification, and that is going to include the operating system. Safety critical certified OS's are not that numerous and they are incredibly expensive. They also require hardware that is going to match the safety critical aspects of failure states, so think higher grade components, connectors etc and redundancy. Very expensive as well, and always behind the performance curve by some margin because time to market for certified hardware is so long.

For the highest level of safety certification used in aerospace, it is possible to see 6 LoC per programmer month. That includes producing verified documentation regarding all the design, requirements, state analysis, validation, verification etc etc. So very expensive to write software as well.

Megabound posted:

No, but I would trust a tuned, trialled and rigorously tested program.

Selecting a emergency landing site would not be the hardest problem in the world. One possible solution would be giving the plane a list of green zones where it can land along with the map data it already has. It already has access to its own location via gps.

Or you could give it a downward facing camera. It knows its elevation and your camera is of known specifications, combined with a edge detection algorithm you could identify large contiguous areas. Give the plane a minimum area to shoot for, and it could take care of the rest.

Like, dude, if it was simple, it would already be done. It just isn't, at all.

Sure, yeah - so you are going to put realtime camera(s) data into your control loops? Maybe use some cool deep learning techniques? Except many image analysis techniques are technically uncertifiable, because they often express invariant solutions or are otherwise fundamentally un-testable. This is a major problem for algorithm design in complex applications in safety critical applications in general, and finding a landing area is just one of many more safety critical capabilities you'd need.

Even if you want to handwave the ridiculously complicated process of reliably determining a safe landing area, you also now need to process a huge amount of data in realtime, and we're still far away from being able to do that without literal tons of computers.


Gromit posted:

I don't have a clue so this is an honest question: what is the current state of software control in modern cars? Do they have control over the engine and braking system such that they could accelerate you down the street and plough into oncoming traffic? If so, loads of us are trusting software with our lives.

Yeah, but here's the thing - the software that processes throttle or brake inputs is pretty simple so the certification for drive by wire systems probably only cost companies a few uh, millions of dollars to do to a reasonable safety standard. Well, actually for toyota it was lot more, after the lawsuits ha ha.

Putting security vulnerable networked devices into same network as the control systems is a whole other ball of wax, but personally makes it clear to me that automakers are not even remotely close to being able to pull off a safe fully autonomous vehicle.

Spatial
Nov 15, 2007

Gromit posted:

I don't have a clue so this is an honest question: what is the current state of software control in modern cars? Do they have control over the engine and braking system such that they could accelerate you down the street and plough into oncoming traffic? If so, loads of us are trusting software with our lives.
They do. Toyoto famously killed several people by doing a terrible job of it. It's such an unbelievably shoddy piece of work it's a good fit for this thread! :v:

That's too vague a consideration though. Drill down a bit. The important thing to ask is why do people trust these control systems in the first place, and how applicable are those reasons to other problems? Well, the degree of trust is a function of their simplicity, predictability, constrained feature set, well defined operating conditions, and the exceedingly high quality standards they must meet.

All those aspects of the system are related. More complexity multiplies - not adds to - the difficulty of making the system work safely and reliably across its various use cases. You need to test most combinations of features and subsystems together or it becomes more and more likely that little Timmy is going to turn into a beef patty due to some oversight. So this burden can very quickly become a profound one.

I work with people who develop these systems and they say just the extra verification work needed to quantify and guarantee safety in their systems adds two extra years to the development time and millions to the cost. These are relatively speaking very simple systems too. Self-driving on the other hand maxes out every single one of those considerations to their craziest levels, multiplying up the verification requirements to a degree that I can hardly even describe. Making something safe despite unlimited complexity, imperfect information, unpredictable conditions, a massive feature set... it's a nightmare project. Making it work at all is the lowest bar possible and even that's a really drat tough one to clear.

Gromit posted:

When it comes to autonomous vehicles, I'll be happy if they perform as good as humans. Performing better would just be icing on the cake. Over 30,000 people die on the roads in the US each year, the bulk of which is a result of driver behaviour. I wonder how many deaths the computers will be allowed before they are considered a failure?
If they're only that good though, it could wind up being too expensive or publically disasterous. It's no longer merely someone dying using our product, but our product failing and killing the customer as a consequence. If that happened often the product's life could also end up coming to an abrupt end.

Also consider that for each fatal crash there are 170 more crashes that wreck the car but don't kill anybody, totalling to more than five million. That's a lot of crashes to be responsible for. Repairs, new cars, hospital costs. All our fault. I don't think we could release a product with that kind of risk without it being the world's biggest albatross around our necks.

Spatial
Nov 15, 2007

drat, I really should be making these long posts at work instead of on Saturday night. SAD!

shame on an IGA
Apr 8, 2005

I work in factory automation and machine building and it's only recently that PLC hardware has been qualified to a level that allows it to control something as simple as the e-stop button and machine guard latches. For the most part, everything safety related is still done through double-redundant hardware relay modules that cost 10x as much as they would without the certification.

packetmantis
Feb 26, 2013
Improbable Lobster is a self-driving car IRL and doesn't want anyone to horn in on his territory.

MrYenko
Jun 18, 2012

#2 isn't ALWAYS bad...

Blitter posted:

Putting security vulnerable networked devices into same network as the control systems is a whole other ball of wax, but personally makes it clear to me that automakers are not even remotely close to being able to pull off a safe fully autonomous vehicle.

Most automakers (I'm looking at you, General Motors,) still seem to treat automotive software as a hard part of the vehicle, and sold under the same warranty. Your Bluetooth integration is so buggy that it flat out doesn't work? Too bad, buy a new car; We don't offer software updates unless it's part of a recall.

This approach to software simply will not continue to work going forward, since as mentioned above, cars are almost all now connected to the internet at some level, and the OE is the author, or at least responsible for, the software in the vehicle. A network vulnerability that allows malicious control of the vehicle is just as dangerous to a nine year old car as it is to a brand new one parked on the showroom floor.

Spatial
Nov 15, 2007

Don't get me started. I've been a part of discussions where department heads were actively pushing to add a permanent hardware backdoor in our security-critical product to reduce risk. The understanding is just not there at a high level in a lot of non-software-specialised companies.

AreWeDrunkYet
Jul 8, 2006

MrYenko posted:

A network vulnerability that allows malicious control of the vehicle is just as dangerous to a nine year old car as it is to a brand new one parked on the showroom floor.

Is there a reason control software can't just be on a separate system than entertainment and other features that could benefit from the connectivity but aren't critical? If the control systems need to be patched, that can be done via physical media at dealerships - not that this can't be hijacked, but it adds a significant barrier and even a vulnerability created this way couldn't be remotely accessed.

karoshi
Nov 4, 2008

"Can somebody mspaint eyes on the steaming packages? TIA" yeah well fuck you too buddy, this is the best you're gonna get. Is this even "work-safe"? Let's find out!

AreWeDrunkYet posted:

Is there a reason control software can't just be on a separate system than entertainment and other features that could benefit from the connectivity but aren't critical? If the control systems need to be patched, that can be done via physical media at dealerships - not that this can't be hijacked, but it adds a significant barrier and even a vulnerability created this way couldn't be remotely accessed.

It would add $10 to the cost of the car, so no, it's not feasible.

Spatial
Nov 15, 2007

karoshi posted:

It would add $10 to the cost of the car, so no, it's not feasible.
This is not a joke. :negative:

Slugworth
Feb 18, 2001

If two grown men can't make a pervert happy for a few minutes in order to watch a film about zombies, then maybe we should all just move to Iran!
This never would have happened if the autonomous road builder had just made the road lower.

shame on an IGA
Apr 8, 2005

karoshi posted:

It would add $10 to the cost of the car, so no, it's not feasible.

more like 3.50 and still unfeasible for the same reason. You think you've seen rear end in a top hat customers working in retail, try selling something to Fiat-Chrysler.

Mycroft Holmes
Mar 26, 2010

by Azathoth
my dad worked at ford in the 90s and these sound like the stories he tells. He wanted to get GPS put in cars, but they wouldn't because they hadn't sold any cars with GPS the previous year. How can you sell cars with GPS when you don't even offer that as an option?

Lime Tonics
Nov 7, 2015

by FactsAreUseless


click for big.

Bloody Hedgehog
Dec 12, 2003

💥💥🤯💥💥
Gotta nuke something

evil_bunnY
Apr 2, 2003

oohhboy posted:

I don't think a computer now or even in many decades time could or would have selected to land on the Hudson or picked the solution in every other emergency that the pilot pulled everyone asses out of the fire. You are asking people to program intuition and to formulate answers in contexts that are well outside it's normal operation with extremely limited information.
Especially because Sully didn't follow the double engine out procedure ( initiated the APU startup out of sequence because he realized he'd be in the drink or worse if he waited until his engines had failed to restart).
Now to be honest less than half the pilots who even knew the solution, and knew how to execute it even managed to land the drat bird without casualties.

Megabound posted:

No, but I would trust a tuned, trialled and rigorously tested program.

Selecting a emergency landing site would not be the hardest problem in the world. One possible solution would be giving the plane a list of green zones where it can land along with the map data it already has. It already has access to its own location via gps.

Or you could give it a downward facing camera. It knows its elevation and your camera is of known specifications, combined with a edge detection algorithm you could identify large contiguous areas. Give the plane a minimum area to shoot for, and it could take care of the rest.
Lmao you have no idea what you're talking about.

Megabound
Oct 20, 2012

evil_bunnY posted:

Lmao you have no idea what you're talking about.

A Vision Based Forced Landing Site Selection System for an Autonomous UAV

One among many of papers exploring using a computer vision system for emergency landings in UAVs.

Blitter posted:

Like, dude, if it was simple, it would already be done. It just isn't, at all.

I never said it was simple, I was exploring a method that could be utilised after a lot of work.

Carbon dioxide
Oct 9, 2012

https://i.imgur.com/OHbjpOU.gifv

PostNouveau
Sep 3, 2011

VY till I die
Grimey Drawer
Isn't the brake on a car just the same physical thing it always was though? Like there might be a program that would allow automatic braking, but it wouldn't ever interfere with the driver's ability to step on the brake and stop the car? I'd heard all those Toyota crashes were likely driver error after it was all said and done.

MrYenko
Jun 18, 2012

#2 isn't ALWAYS bad...

PostNouveau posted:

Isn't the brake on a car just the same physical thing it always was though? Like there might be a program that would allow automatic braking, but it wouldn't ever interfere with the driver's ability to step on the brake and stop the car? I'd heard all those Toyota crashes were likely driver error after it was all said and done.

Brakes generally still have a physical connection between the master cylinder and the pedal. The throttle no longer does; It is completely electronic.

Adbot
ADBOT LOVES YOU

PostNouveau
Sep 3, 2011

VY till I die
Grimey Drawer

MrYenko posted:

Brakes generally still have a physical connection between the master cylinder and the pedal. The throttle no longer does; It is completely electronic.

Alright, so all this "Macedonian teenagers are going to hijack your Internet-connected car on the freeway and hold you ransom for bitcoins" stuff is malarkey. There's not a car in the world that will keep going when the brake is engaged.

  • 1
  • 2
  • 3
  • 4
  • 5
  • Post
  • Reply