alt Hacker News> The kit utilizes a novel interface with the aircraft’s flight controls and mission systems, allowing a pilot to toggle between traditional human control and AI control with the flip of a switch. This ensures a safe, reliable environment for human-on-the-loop experimentation.
How, exactly, does that ensure a safe reliable environment? Humans are pretty bad at suddenly taking over when an automatic system reaches its limits and throws a problem in the human's lap. A lot of air upsets and crashes start when the autopilot disengages and hands control of a badly configured aircraft back to a human pilot.
Replies
These are going to be experienced test pilots who are pretty good at actively managing autopilots and other automated systems. Plus they'll be backed up by telemetry to ground controllers. But ultimately the military is willing to accept some losses in order to advance key R&D programs.
The Tesla method! Disengage shortly before a crash. Now the human is responsible for it.
It doesn't, flying these is potentially very dangerous, there are many non-recoverable situations. But AFRL/NASA has been working on this for decades. The manoeuvres the higher level AI requests are actioned by highly assured dynamical systems solvers using model predictive control, like the autopilot, that maintain the aircraft in the safe envelope.
The unknowns in this system are (a) how agressive is the flight controller? (b) how clever is the autonomy?
Probably a duplicate autopilot-like system that an AI controls. Flip a switch and controls revert back to the pilot. It may not even be a real switch, many planes disengage AP when you touch the controls.
I doubt AI will be doing dogfighting anytime soon, this is likely just an AI following predetermined routes and maneuvers and making the choice on how to accomplish them. The great thing about computers is that they can process inputs faster than humans.
Once the AI realizes it can do a maneuver the aircraft can survive but the "bossy meatbag" in the cockpit cannot...
F-16's CFIT reminds me of Boeing's Controlled Impact Demo in the 80s out of Edwards AFB: https://en.wikipedia.org/wiki/Controlled_Impact_Demonstratio...
Thats a standard cookie cutter statement intended to give an impression of safety.
It's "human-on-the-loop experimentation", not mass production. Are you engineer? How do you experiment? I assume by pushing the boundaries step by step? How do you suggest they do testing if not by taking risk?
The switch is not novel, that is in fact exactly how autopilot works for any aircraft. The pilot toggles it on and off and the autopilot itself may turn itself off if certain situations occur.
To your point, that isn't without problems, but it certainly isn't novel. It is very standard and boring.
I would bet this is pure marketing speak and there really isn't anything special here. "Follow this flight plan and do so within this flight envelope" is pretty straightforward?
Flying is not driving. Things dont happen as quickly. Kids dont step out in front of jets at 30,000 feet. Air is also homogenous. Pavement is not. Whereas a car must turn or die, a jet can generally fly straight and level without issue while problems get solved. An out-of-control car has maybe three seconds before a collision. A jet will generally have over a minute to get back under control.
Sure, there are moments. Landing and takeoff are very time sensitive operations, but those also occur in protected environments, sanitized of the random issues that cars must deal with.
And the AI isnt doing all the flying. The aircraft has many protective systems to assist the pilot regardless of who that pilot may be. (Stable aerodynamics, anti-stall limits, control lockouts etc.)
This stuff actually gets tested, it's not your typical web framework junk.
A computer is "flying" the F-16 all the time anyway. It's a statically unstable aircraft.
I wonder if the AI control is giving inputs to the same flight controls system a human pilot would, or if it has a lower-level interface that could bypass the control laws in the human-operated system.
On one hand you might be interested in the AI coming up with ways to fly the aircraft that the normal control laws don't allow, but on the other hand you don't want the AI coming up with exciting new ways to depart from controlled flight (potentially in unrecoverable attitudes) with a pilot in the very expensive aircraft (even if the pilot could eject, which is an endeavor fraught with risk to life and limb).
Edit:
Presumably since there's a pilot in the aircraft the AI must be bound by at least some of the regular control laws that prevent the pilot from doing maneuvers that would, from a G-loading perspective, be incompatible with human life.
At least one would hope.
I don't think I'd go flying with a virtual pilot that was capable of flying the aircraft in a manner that could kill me but otherwise leave the aircraft intact.