Go Back  PPRuNe Forums > Flight Deck Forums > Tech Log
Reload this Page >

SOP FOCM etc driven (hypothetical) Artificial intelligence pilot.

Tech Log The very best in practical technical discussion on the web

SOP FOCM etc driven (hypothetical) Artificial intelligence pilot.

Reply

Old 15th Apr 2018, 23:10
  #1 (permalink)  
Thread Starter
 
Join Date: May 2010
Location: Boston
Age: 67
Posts: 236
SOP FCOM etc driven (hypothetical) Artificial intelligence pilot.

The recently revived 'A320 can you do auto-land overweight?' thread made me remember this question I have been pondering for a while:

Given an (obviously hypothetical at present state of the art) AI 'pilot' capable of reading and understanding all available documents, including the break rules as needed in case of emergency "prime directive".

A few questions to ponder:

Any examples of real world scenarios that a deeply experienced pilot would successfully handle that the AI 'pilot' might not?

Note "real world" in above, possibly counterintuitively, anything found in standard simulator training, even the most extreme cases would be handled perfectly since the "AI pilot" would have been trained.

How does this compare to a freshly minted minimal hours pilot?

Should/could anything change in current docs based on answers to above?

Lastly, would automatic discovery of incomplete or conflicting documentation be useful?

Note. I am not a pilot but have a deep personal interest in human factors and professional experience in automatic verification of specifications.
Just noticed FOCM/FCOM typo in title as soon as I submitted, don't think I can change that.

Last edited by MurphyWasRight; 16th Apr 2018 at 14:52. Reason: Corrected typo in title
MurphyWasRight is offline  
Reply With Quote
Old 15th Apr 2018, 23:32
  #2 (permalink)  
 
Join Date: Dec 2006
Location: The No Trangression Zone
Posts: 2,049
The MD 11 can be trained to handle an Al Haynes scenario. This is called "PCA" (Propulsion Controlled Airplane)

However, I highly doubt AI would think to cut out the anti skid so that the tires burst in a 737 after the ramp worker left a chock in the gear well leading to the landing gear failing to extend
Pugilistic Animus is offline  
Reply With Quote
Old 16th Apr 2018, 01:16
  #3 (permalink)  
 
Join Date: Dec 2006
Location: The No Trangression Zone
Posts: 2,049
My apologies I meant one LDG gear failed to extend hence there was assymetry I'll look for the video
Edit Left Main Gear it was
Let's go to the videotape
Pugilistic Animus is offline  
Reply With Quote
Old 16th Apr 2018, 01:43
  #4 (permalink)  
Thread Starter
 
Join Date: May 2010
Location: Boston
Age: 67
Posts: 236
Thanks, that is the type of "real world" problem I had in mind.

Something not be covered by any conceivable SOP that used operation outside limits to cause deliberate a deliberate failure to improve the overall chances.

Any thoughts on what a low experience "just follow the rules" pilot would have done?

True AI would be able to synthesis novel solutions from available data and has been "in about 10 years" for the last 30 or more.
MurphyWasRight is offline  
Reply With Quote
Old 16th Apr 2018, 04:13
  #5 (permalink)  
 
Join Date: Dec 2006
Location: The No Trangression Zone
Posts: 2,049
Low experienced pilot? I don't have a clue what they might do...depends on training history and mindset, I guess
Pugilistic Animus is offline  
Reply With Quote
Old 16th Apr 2018, 04:17
  #6 (permalink)  
 
Join Date: Jul 2006
Location: Somewhere over the rainbow
Posts: 382
Originally Posted by MurphyWasRight View Post

Any examples of real world scenarios that a deeply experienced pilot would successfully handle that the AI 'pilot' might not?
Landing in the Hudson river case?

You can correct your title if you go on “Edit” then “Go Advanced”.
pineteam is offline  
Reply With Quote
Old 16th Apr 2018, 04:49
  #7 (permalink)  
 
Join Date: Dec 2006
Location: The No Trangression Zone
Posts: 2,049
I think any scenario where gliding or crash landing will really over tax AI...and I mean engine failure even on a Cherokee is a very difficult computational problem, actually...I'm a non expert in computers though...in fact I'm such a NeoLuddite that I still wish that you go to the nearest computing office and have the girls draw all you're charts and graphs

Man has really come a long way
Pugilistic Animus is offline  
Reply With Quote
Old 16th Apr 2018, 15:04
  #8 (permalink)  
Thread Starter
 
Join Date: May 2010
Location: Boston
Age: 67
Posts: 236
pinetime: Thank, I corrected title. I had originally missed that I needed to edit the original post.

On landing power off or in Hudson etc:

The control of the aircraft with no power or multiple system failures is well within capabilities of today's systems, although not widely implemented since the current approach to automation is to let the pilots handle the unexpected.

The wisdom of that is a whole other thread, see AF477.


What is a more difficult task would be the decision to land in Hudson.

I believe that is where deep experience saved the day.
A low hours pilot would possibly gone for the Teterboro option with probably 'sub optimal' results.
MurphyWasRight is offline  
Reply With Quote
Old 16th Apr 2018, 15:31
  #9 (permalink)  
 
Join Date: May 2017
Location: Hawaii
Posts: 38
I never understand why the Hudson river is used as an example of amazing/difficult decision making.

Loss of thrust in both engines means you will be touching the ground/water in about as many nautical miles as twice your height in 1,000s of feet. So for the Hudson river case about six miles. Were they within six miles of a runway? No. So it will be an off runway ditching/forced landing. Anyone who takes off over a large city or hazardous terrain will/should have thought about what they'd do in case of a forced landing at various heights. Hell, we even give it a whirl in the sim for a bit of fun at the end of the session.

Over a massive conurbation with a long and wide river running through it which option do you take? Crashing into a city or ditch in the river? Again, this is not a tricky question.

As for actually flying the Airbus into the water; Sully had full backstick maintained all the way down below about 100ft and as a result could not flare to reduce the descent rate at impact. See section 2.3.3 of the NTSB report. In essence, normal law of the Airbus did what it was designed to do - max alpha until impact.

Are you saying a computer could not be programmed with some basic rules to perform similarly and perhaps better? And I am not denigrating or taking away from what Sully and his crew did on that day whatsoever.
Its Maui is offline  
Reply With Quote
Old 16th Apr 2018, 15:34
  #10 (permalink)  
 
Join Date: May 2017
Location: Hawaii
Posts: 38
As for the 737 incident above, it worked out for him on the day. Is it procedure to do that for a main gear up landing on the 737? It isn't on the A320. So I would hazard a guess that the manufacturer and test pilots etc put their heads together and decided it's not the best option to do what he did. So in that regard, was it necessary and can computer be programmed to think of it as an option?
Its Maui is offline  
Reply With Quote
Old 16th Apr 2018, 15:57
  #11 (permalink)  
 
Join Date: Dec 2006
Location: The No Trangression Zone
Posts: 2,049
It's quite easy to see how far you're gonna get...Forget turning LGA and TEB!
Pugilistic Animus is offline  
Reply With Quote
Old 16th Apr 2018, 16:22
  #12 (permalink)  
 
Join Date: Dec 2006
Location: The No Trangression Zone
Posts: 2,049
That 737 pilot was for his flight a test pilot...so we're Al Haynes and Denny Fitch so was Sully
Pugilistic Animus is offline  
Reply With Quote
Old 16th Apr 2018, 17:45
  #13 (permalink)  
 
Join Date: May 2017
Location: Hawaii
Posts: 38
And they all did great jobs. But it's still interesting that in the gear stuck up incident his actions were not then incorporated into checklists and actions for future events.
Its Maui is offline  
Reply With Quote
Old 16th Apr 2018, 18:19
  #14 (permalink)  
 
Join Date: Dec 2006
Location: The No Trangression Zone
Posts: 2,049
Too rare...
Pugilistic Animus is offline  
Reply With Quote
Old 16th Apr 2018, 18:20
  #15 (permalink)  
 
Join Date: Jul 2006
Location: Somewhere over the rainbow
Posts: 382
Originally Posted by Its Maui View Post

Are you saying a computer could not be programmed with some basic rules to perform similarly and perhaps better? And I am not denigrating or taking away from what Sully and his crew did on that day whatsoever.

I do agree the ditching part is not that impressive but I don’t beleive IA, at least nowadays would be able to take that kind of split decision. Even if it was smart enough to decide to ditch, it might crash into a ferry passing by.

Last edited by pineteam; 16th Apr 2018 at 18:32. Reason: Typo.
pineteam is offline  
Reply With Quote
Old 16th Apr 2018, 20:50
  #16 (permalink)  
Thread Starter
 
Join Date: May 2010
Location: Boston
Age: 67
Posts: 236
I do agree the ditching part is not that impressive but I donít beleive IA, at least nowadays would be able to take that kind of split decision. Even if it was smart enough to decide to ditch, it might crash into a ferry passing by.
I don't remember if it was possible to see and actively avoid obstacles on the Hudson or whether the "big river" effect was in play.

As others have pointed out once all engines quit you will soon be landing somewhere.

Actually not that hard to have the AI handle that scenario by locating the "most suitable location" following a simple enough tree:

Inside predicted range including obstacles etc:

Suitable airport
Any airport, even a short runway likely better than most "other"
smoothest other surface, around NY that would most likely be water.
avoid school yards. ( As the newspapers like to report.)

Interestingly this would avoid the sometimes seen private pilot unsuccessful attempts to return to take off RWY on engine failure when the corn field is a safer option.

It's quite easy to see how far you're gonna get...Forget turning LGA and TEB!
I remember (possibly incorrectly) that TEB was offered by controllers, I do remember someone saying it "might" have been theoretically possible if everything went perfectly.
MurphyWasRight is offline  
Reply With Quote
Old 16th Apr 2018, 21:26
  #17 (permalink)  
 
Join Date: May 2017
Location: Hawaii
Posts: 38
Exactly. This is not a split second decision. It's something that responsible pilots should have considered well before a bad day happens. In fact; arm-chair flying through random, crazy inflight emergencies is very helpful for these kind of situations.

And if we can consider it calmly and come to solid conclusions then it is quite easily programmable.

Adding crashing into a ferry after a loss of thrust in both engines is just your day to die. Sorry but what if it had been total IMC? What if both wings fell off?
Its Maui is offline  
Reply With Quote
Old 16th Apr 2018, 21:37
  #18 (permalink)  
 
Join Date: Jul 2013
Location: Everett, WA
Age: 63
Posts: 1,817
AI would be potentially useful for 'it's never happened before' scenarios (which is where humans can often excel), but part of the process for implementing a 'pilotless' aircraft would be to list every single emergency scenario that's ever happened (at least the ones that are potentially survivable - the wings falling off can probably be excluded). Then determine the optimal course(s) of action and program that into the computer. An all engine power loss is actually one of the easier scenarios - it's happened many times so there is considerable data available as to what course of action would be most successful. The primary advantage of the computer is it would be able to analyze all the potential options and chose the most viable course in a fraction of a second. Sully did fantastic, but it took him several seconds to make that decision which unfortunately eliminated potential options.
tdracer is offline  
Reply With Quote
Old 16th Apr 2018, 22:02
  #19 (permalink)  
 
Join Date: Dec 2006
Location: The No Trangression Zone
Posts: 2,049
Yes indeed "The impossible turn" applies to all planes...you have to just turn your head foward and never look back again
Pugilistic Animus is offline  
Reply With Quote
Old 16th Apr 2018, 22:33
  #20 (permalink)  
Thread Starter
 
Join Date: May 2010
Location: Boston
Age: 67
Posts: 236
part of the process for implementing a 'pilotless' aircraft would be to list every single emergency scenario that's ever happened (at least the ones that are potentially survivable - the wings falling off can probably be excluded). Then determine the optimal course(s) of action and program that into the computer
Keeping in mind this is a hypothetical AI system that is not yet available:

The list of emergency scenarios would be better used as a part of a test program.

The AI pilot should be "taught the basics" and have access to as much type specific data as possible including SOPs and other docs.

The database used for flight simulators would be a good addition since it would allow modelling outcomes of multiple possible actions. (This one is probably a few years out in terms of cpu resources and response time)

Given this and some relatively broad "prime directives" the AI pilot should be able to derive the optimal action when presented with a novel (to it) situation.

If it fails then analysis of the failure and further training would be indicated.


Anybody know if a flight simulator would correctly model the one gear down incident?
MurphyWasRight is offline  
Reply With Quote

Thread Tools
Search this Thread

Contact Us Archive Advertising Cookie Policy Privacy Statement Terms of Service