OK, how would the computer handle this one, which actually happened. FO flying, A/P 2 engaged. Cleared to FL230, which the aircraft merrily sailed through. FO levels and returns to FL230. re-engages A/P 2 and aircraft starts to climb.
Trivial. Triple (or more) redundant systems with voting. Herod, you seem to be latched into pilotless airplanes using the same systems that are in use today - systems that were intentionally designed for human input/interaction. Autonomous systems simply wouldn't work that way.
And you seem to be hung up with the ability of a computer to 'see'. Except that humans can only really see well when it's clear and daylight or lit - the rest of the time we depend on, wait for it, computer generated radar and night vision
TCAS 'sees' other aircraft well beyond human visual range, figures out what to do about it, and tells the pilots. Since the widespread adoption of TCAS, there has been one major mid-air, and that occurred when the pilots ignored what TCAS told them and did what a human controller told them instead
Most of the automated systems on todays airplanes are archaic and horribly outdated because it's so expensive to certify updates.
BTW, since Sioux City was brought up, another minor detail. Yes, human pilots did a remarkable job of flying that airplane - I understand that when they put the scenario in a simulator most pilots crashed within minutes. When the idea of training pilots to be able to fly an airplane without hydraulics by using differential thrust, it was rejected as impractical - that scenario was so rare that it would be waste of simulator time. Yet, in the aftermath, there was a team at Boeing working on a computer program that would do exactly that for the 777 (they focused on the 777 as it was the only Boeing product at the time where the throttles could be independently controlled automatically - today 787 has that same capability). Now, I don't know what ever became of that study (I moved off the 777 program shortly afterward and lost visibility), but if they could get it right, it would always be right - unlike humans you can transfer complete and total knowledge between computers.
Total loss of hydraulics is just one example - there are many other examples of failures that are considered so remote that training is not justified, yet we've had to redesign system hardware and software to accommodate or preclude the extremely rare failure - in some cases simply in order to certify - since we couldn't take credit for crew action as they never train for it
Oh, and someone brought up the first moon landing. You do realize that the LEM had less computing power than my $20 digital watch? Oh, and 7 years later we autonomously landed two Viking landers on Mars?
