DARPA human vs AI BFM trials

Started by Spam, August 20, 2020, 05:04:32 PM

0 Members and 1 Guest are viewing this topic.

Spam

As a working defense engineer and with a background in 5th gen fighter design/flight test, I don't want to tempt Cadets to skip school, but live streaming right now is the DARPA series of flight tests in which AIs are competing and ultimately will "fight" a human pilot, engaging in mock BFM combat.  So the AI doesn't get tired, and the AI is learning..  how will it end?



Spam

Results (Spam's take)

5 - 0 to the AI.

From five neutral BFM sets, the experienced human pilot (a USAF Weapons School grad) simply was unable to beat the AI with (simulated) guns using conventional tactics. Factors to note include the literally superhuman ability of the Heron AI to perform continuously computed WEZ (weapon engagement zone) calculations real time in turning and in head on passes, to support consistently accurate snap shots against fast fleeting targets. This, where a human has to perform an estimated snap shot, while an AI can run massive calculations, playing to its strength there. Also, in this simulated environment, the human pilot was able to survive (having learned some TTPs that worked against the AI) but to do so, he pulled over nine Gs continuously for several minutes in scissors reversals and a continual turning fight; few humans would be able to maintain the required accurate touch on the stick while maintaining that G load, where an AI could care less about Gs.

Caveats:  modeling as always has inaccuracies (aero models, weapons models, etc.), training for the pilot may be a factor (new video game style PVI/not representative of an actual Viper), etc.  Also, WVR (within visual range) BFM (basic fighter maneuvering - gun fighting) for modern 5th gen aircraft is playing to our weaknesses, not strengths; a properly trained 5th gen pilot will fight unfairly, using sensors, datalinks, observability and speed, to engage targets from BVR, not in a knife fight. It will be very interesting to see this program move to a 2v2 or v many evaluation, and to see the agents used to evaluate BVR engagements, perhaps to support the nascent Loyal Wingman program elements.

Hopefully DARPA will leave this up as an archived presentation for interested members to review; as an unclassified sim event I found this accessible to cadets, technically worthwhile to watch, and easily comparable in terms of engineering quality with similar fighter events I've sponsored and run over the years. Kudos to Johns Hopkins APL, and all the AI devs.

V/r
Spam

TheSkyHornet

I don't know how I'm going to find 5 hours to watch this, but I really want to nerd out right now.

This is just so friggin' cool.

PHall

Now, was the AI being tested in a computer that is small enough to be in a "fighter" sized aircraft?
Or was this a big "super computer"?

And how does it get past the First Law of Robotics? ;)

Spam

Quote from: PHall on August 20, 2020, 08:45:55 PMNow, was the AI being tested in a computer that is small enough to be in a "fighter" sized aircraft?
Or was this a big "super computer"?

And how does it get past the First Law of Robotics? ;)

As to the first question, the CIPs on the F-22 fit quite well. They were so freakin hot, however, they required extraordinary measures to cool them (PAO fluid, only the second platform I've worked on which used cold plates, the other being the space shuttle). If interested, a former coworker published on that at:  https://www.jstor.org/stable/44723138?seq=1, and there is a nice article on the Increment 3.2 package at: https://www.aviationtoday.com/2018/10/11/rejuvenating-raptor-roadmap-f-22-modernization/

As to the second, the process is the same as training a young E-1: aim at a target, and don't think of it as a person, I suppose. Susan Calvin was to have been born in 1982, so she should be finished with her PhD by now and be working in the field... perhaps she is working on one of the teams!

Spam

Funny... sat down with dinner, and AMC channel is playing "I, Robot" right now.

;D

coudano

Quote from: Spam on August 20, 2020, 09:37:20 PMThey were so freakin hot, however, they required extraordinary measures to cool them (PAO fluid, only the second platform I've worked on which used cold plates, the other being the space shuttle).

So maybe the AI *does* care about sustained G's ;)