How close are we to commercial aircraft with autopilots that can, as safely as theoretically possible, land an aircraft with damaged, missing, or otherwise inoperative control surfaces; or altered aerodynamics (e.g. ice on the wings), missing or misleading avionics (e.g., frozen pitot tubes), etc.?

Remember the Gimli Glider? An autopilot that could do that would be a good start.

And the computer in Space Shuttle Columbia made a valiant effort to keep the nose pointed in the right direction even as the aircraft was disintegrating.

  • $\begingroup$ Generally the autopilot functions rely on sensors and actuators/control surface intact, and cease to function when this is not the case. When a pitot is frozen, there are inconsistencies in pressure data (with redundant systems) and automation hand over control to the crew. The scenario of Gimli glider can't be entirely delt by current automation (starting with a first problem: the airfield was decommissioned and not in FMS database). $\endgroup$
    – mins
    Dec 6, 2017 at 8:13
  • $\begingroup$ To expect a computer to analyse context and adapt no new situations is the same to expect to a human to react faster and with more precision than a computer. We are better than computers in some areas and computers are better than us in some aspects, let's use computers to suplement our abilities and surpass ours limitations. Trying to use it to fully replace us is foolish out of scifi. $\endgroup$
    – jean
    Dec 11, 2017 at 11:19
  • $\begingroup$ This is really a two step job: first is to recover control authority with alternative controls, then second is to evaluate the result of the recovery and land the aircraft within the limits. Computer is good at the first but humans is better at the second, currently. Like if you lose the vertical stabilizer and rudder completely, it's almost impossible for a human pilot to maintain stability and control with differential throttle and air braking, but computer can do it OK. After that is handled, the pilot can just land the plane "normally". $\endgroup$ Dec 11, 2017 at 15:46

1 Answer 1


Reconfigurable flight control systems have been a topic of research in military aircraft for decades. The science is existing and mature, but not an easy one: damage cannot be predefined, and therefore cannot be explicitly programmed in.

Classical re-configurable flight control systems attempt to re-program on the fly: the transfer function between input and expected aircraft response is constantly monitored, and if actual response differs from expected response, the gains in the loop are reset on the fly in order to get closer to a stable, matching response with neutral trim.

From this paper, which is almost 20 years old:

The goal of control reconfiguration is to maintain handling qualities in the presence of a large universe of damage and failure modes. In contrast to robust control design methods of the past, the emphasis in control reconfiguration involves a combination of on-line parameter identification control redesign and/or adaptation for a degraded mode of flight.

Neural networks are now being implemented in reconfigurable flight control systems. These work as the synapses in our brain: there are a lot of possible connections, which are defined in our childhood as a result of positive and negative experiences. The system is self learning, from a multitude of outcomes that are labeled either "desirable" or "not desirable". Neural networking is what makes Google Translate possible; translating a language by programming is just not possible because of all of the different inflictions and double meanings in a language.

So in military aircraft an auto-damage-pilot is very well possible. Civil airliners are a different matter, not because of technology but because of certification. Safety is of the highest importance, and how can we prove that our re-configuration is actually helpful?

  • 1
    $\begingroup$ Neural networks are not needed. There are controllers that are damage-resistant and do not use them. And they will be employed on civilian aircraft. $\endgroup$
    – Federico
    Dec 6, 2017 at 9:43
  • 1
    $\begingroup$ @Koyovis How to prove the AI is safer than human pilots in the real world? Whenever an aircraft experiences a failure that could (or did) bring down the plane, model it in a simulator and test that scenario on the AI and on human pilots. $\endgroup$ Dec 6, 2017 at 17:14
  • 1
    $\begingroup$ Yes a simulator helps - problem is that normal training simulators have a data pack that stays within normal flight conditions, AoA about +/- 20 deg and angle of sideslip about 30 deg. How would you get data of a damaged aircraft to validate the simulator against? And how to implement higher AoA and sideslip data? CFD, wind tunnel data etc is a possibility, but it is still computed data, not real life flight data and that is what the authorities want to see. $\endgroup$
    – Koyovis
    Dec 7, 2017 at 9:45
  • 1
    $\begingroup$ @Koyovis why would the authorities want to see real life flight data? $\endgroup$
    – user7241
    Dec 9, 2017 at 21:14
  • 2
    $\begingroup$ @Antzi It is not difficult to modify existing aircraft dynamics models to account for damage. One just has to change several parameters. Also there have been flight tests conducted with small-scale models whose wing was removed during flight. In principle one could use real aircraft, but this should be expensive, because existing remote controlled drones like the QF-16 might lack the necessary computing power for running resilient flight control systems. $\endgroup$
    – user7241
    Dec 12, 2017 at 4:07

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.