General Forums => The O' Club => Topic started by: FLS on August 23, 2020, 01:49:21 PM
Title: DARPA AlphaDogfight Trials Video
Post by: FLS on August 23, 2020, 01:49:21 PM
The AI competition with the AI winner vs Human pilot at the end. It's a gunfight so the maneuvering will be familiar.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Vulcan on August 23, 2020, 07:19:31 PM
Huh what VR headset is that? I think it may be one of the Oculus half-dome prototypes (it's the only one I can find with a circular back strap like that)
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Ack-Ack on August 24, 2020, 12:03:55 AM
I watched the entire trials, the most common merge the AI pilots maneuvered for was the head on, especially the more aggressive AI pilots.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Eagler on August 24, 2020, 06:20:41 AM
I watched the entire trials, the most common merge the AI pilots maneuvered for was the head on, especially the more aggressive AI pilots.
Isn't that just coding?
They are as smart as they have been programmed.
Watching them fly in endless circles trying to get behind each other looks like additional manuvers could be added
Eagler
Title: Re: DARPA AlphaDogfight Trials Video
Post by: FLS on August 24, 2020, 11:02:48 AM
Giving the bandit a bad shot in exchange for improved position may not be a good strategy with AI shooters.
The AI didn't maneuver for the head on shot, they both maneuvered for a shot and the head on was the result.
When it's AI fighting, the high aspect shot is literally the logical outcome of a neutral start with equal aircraft performance and pilot performance. The winning AI was the better shooter.
The human pilot doesn't even train for that shot. Due to safety concerns he gives turning room to the bandit, and the bandit reciprocates, when training in real aircraft.
The trials are a work in progress and I didn't see any collisions so I assume they were turned off and future development will include collision avoidance with aircraft and debris which would change the gunfights a little.
It's interesting how AI that's programmed to learn from it's mistakes ends up flying textbook maneuvers. I'd like to see 2v2 now.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Vulcan on August 24, 2020, 04:45:31 PM
The first team that puts a turret on their platform will win.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Wolfala on August 25, 2020, 07:41:15 AM
Paging 999000
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Wolfala on August 25, 2020, 07:46:27 AM
To be honest I'd like them to take some top tier sim pilots and put them up against that AI.
Exactly. Take some guy with 20+ years of experience between falcon DCS or aces high and I suspect probably correctly that the AI is not going to stand much of a chance
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Wiley on August 25, 2020, 01:23:17 PM
Just the gunnery alone would make it difficult for the human, but I would be curious.
The AH guy would probably get HOed down and then whine about it for half an hour.
Wiley.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Eagler on August 25, 2020, 02:14:09 PM
They need to factor in g forces for AI or what is the sense
Eagler
Title: Re: DARPA AlphaDogfight Trials Video
Post by: guncrasher on August 26, 2020, 01:51:58 AM
wasn't supposed to be a fair simulation fight. it was meant to test ai.
for those who think if you play simulators for fun and they could be better than a trained pilot, well :rofl :rofl :rofl
semp
As noted in the video a trained pilot has to unlearn a whole pile of stuff that becomes irrelevant or even an obstacle in a sim, such as use of negative G manoeuvres. Being a simulation many of the nuances of operating the aircraft a pilot is trained in simply go away, for example they used air starts.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: FLS on August 26, 2020, 09:55:38 AM
We have real fighter pilots and instructors in AH. They don't automatically dominate but they get better faster than untrained pilots because they have less to learn in the game. They also have qualities that real fighter pilots need that aren't required or useful in the game.
The biggest advantage the AI had in the video was perfect SA. The pilot had to look around with a VR headset and likely couldn't monitor speed while tracking the bandit, I doubt he had a virtual helmet HUD. The AI could optimize turn rate and maneuvering for position. This is why the AI often ended up nose to nose and the best shooter won.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Eagler on August 26, 2020, 10:16:39 AM
How much longer is a human fighter pilot even needed in the cockpit?
Eagler
Title: Re: DARPA AlphaDogfight Trials Video
Post by: DmonSlyr on August 26, 2020, 11:58:09 AM
How much longer is a human fighter pilot even needed in the cockpit?
Eagler
Probably not much longer TBH, unfortunately.
Check out a company called Kratos Defense. (KTOS). Its right around the corner.
As stated, the biggest advantage are views for SA and G Forces, also not risking a humans life.
I find it interesting that there were no HOs after the first merge. Even the AI has respect for a real fight. How about that.
I don't agree with AI only fighters in real life. We don't need a real Skynet, even though there is... we should always have a real human behind the stick.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: FLS on August 26, 2020, 12:23:39 PM
If you watch all the fights you'll see that they take every HO shot opportunity.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: DmonSlyr on August 26, 2020, 12:30:34 PM
If you watch all the fights you'll see that they take every HO shot opportunity.
Okay after watching it more I realize they were shooting in the HO. Didnt catch it the first time. Thats kind of too bad because I feel like HOing in these situations defeats the purpose.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Wiley on August 26, 2020, 01:10:15 PM
Okay after watching it more I realize they were shooting in the HO. Didnt catch it the first time. Thats kind of too bad because I feel like HOing in these situations defeats the purpose.
Now there's an interesting question... Especially in a sim, you *should* be able to have an AI have near perfect gunnery... Could you make an AI that could successfully avoid the HO or would the gunnery always win out? That would be an interesting test/challenge.
Wiley.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Haskell on August 26, 2020, 01:20:53 PM
Watching them fly in endless circles trying to get behind each other looks like additional manuvers could be added
Eagler
endless circles are the Nash Equillibrium, if they deviated they would be putting themselves at a disadvantage. remember that most pilots make tons of mistakes from their POV, theres an element of deceit and trickery involved in human v human.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: FLS on August 26, 2020, 02:54:56 PM
Okay after watching it more I realize they were shooting in the HO. Didnt catch it the first time. Thats kind of too bad because I feel like HOing in these situations defeats the purpose.
You can see the damage meter go down when they get hits.
Their purpose is maneuver to take a shot. You see this in AH too, when both maneuver well you end up nose to nose. You believe in exceptions to taking the shot but that's a gaming culture belief. In combat the worry is not getting the shot first. That's not to say you'd choose a HO if you have a choice but the AI fights show the logical outcome of equal performance.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Vulcan on August 26, 2020, 03:06:28 PM
All the talk about maneuvering is pointless when a simple turret will end the argument. You have to stop thinking traditionally.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Wiley on August 26, 2020, 03:24:22 PM
All the talk about maneuvering is pointless when a simple turret will end the argument. You have to stop thinking traditionally.
Well, if you're going for effectiveness, a swarm of smaller cheap turreted drones would be the way to go. Harder to kill, more options for destroying the enemy.
But that's boring. ;)
Wiley.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: DmonSlyr on August 26, 2020, 07:26:18 PM
You can see the damage meter go down when they get hits.
Their purpose is maneuver to take a shot. You see this in AH too, when both maneuver well you end up nose to nose. You believe in exceptions to taking the shot but that's a gaming culture belief. In combat the worry is not getting the shot first. That's not to say you'd choose a HO if you have a choice but the AI fights show the logical outcome of equal performance.
Wellll.... I believe avoiding the HO shot and getting the angle before the opponent if he tries to HO shoot is a lot less risky and more beneficial. I feel like HOing after the first pass defeats the point of the duel. As Wiley noted it would seem like the AI should have the better aim given its computer. In real life I would not risk going to the HO shot in a duel because it seems like a lose lose for both parties. That's normally what happens to me in AH.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: guncrasher on August 26, 2020, 07:57:21 PM
Wellll.... I believe avoiding the HO shot and getting the angle before the opponent if he tries to HO shoot is a lot less risky and more beneficial. I feel like HOing after the first pass defeats the point of the duel. As Wiley noted it would seem like the AI should have the better aim given its computer. In real life I would not risk going to the HO shot in a duel because it seems like a lose lose for both parties. That's normally what happens to me in AH.
ai never got hit, so HO shots worked for it.
semp
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Eagler on August 27, 2020, 07:32:52 AM
If AI is not blacking out due to G forces, wouldn't it always win as it will be able to turn sharper than a passing out human pilot?
Turn fights in jets seem so 1950
Pretty sure with todays tech it rarely/never happens in real combat
Eagler
Title: Re: DARPA AlphaDogfight Trials Video
Post by: FLS on August 27, 2020, 09:06:11 AM
All the talk about maneuvering is pointless when a simple turret will end the argument. You have to stop thinking traditionally.
The real purpose isn't trying to win gunfights, they're demonstrating what AI can do within the simple confines of a traditional 1v1 similar aircraft gunfight. The first goal is building trust in AI systems that can enhance human pilots.
Title: Re: DARPA AlphaDogfight Trials Video
Post by: Eagler on August 28, 2020, 06:02:04 AM
Issue with AI is that it can be hacked
See US drone captured by Iran
Eagler
Title: Re: DARPA AlphaDogfight Trials Video
Post by: FLS on August 30, 2020, 12:23:09 PM
Heron explains how they won. One point to understand about the trials is that the AI is teaching itself to fight, it's not following specific maneuvers programmed by people.
Another point is that there was no shoot now decision, if the bandit was in the gun cone it took damage.
Also, you might notice in the first video above that the aircraft are shown hugely oversize so that the aircraft attitudes are obvious in the videos when zoomed out enough to see the turn circles, this makes the turn circles look tiny and the speeds look slow in the video.