AI bug

[Predator555 0]

i am responding to a thread i created on the armaholic forums about an AI bug, the thread can be found here

http://www.armaholic.com/forums.php?m=posts&q=3272

this issue was raised on these forums too by a kind member of the armaholic forums who was willing to post it because i could not log in on these forums for some unknown reason, and there was some discussion on the matter but for some unknown reason that thread was locked

the thread can be found here and the relevant post is by Alex[Dev]72

http://www.flashpoint1985.com/cgi-bin....1247879

anyway i managed to gain access to these forums once again and can now post, the issue is as described in the previous threads, when setting up an ambush groups of AI have unrealistic abilities to detect you

for example when setting up an ambush at close to medium range if a single soldier walks past as expected you go unoticed and he walks straight past, however if it is a 3 man patrol that is walking past you without ever looking at you they detect your presence. i have a video that i will add to this once it has uploaded to illustrate it a little better.

[KeyCat 131]

As mentioned in the above thread I tried to reproduce this several times with the 1.12 Beta but couldn't see anything wrong!? If I stayed well hidden the group of AI could pass me by within 2 meters and I was undetected.

If you can, please share a simple test mission and a short description how to reproduce it...

/KC

[Predator555 0]

im uploading a video as we speak but yes i will try to create a test mission

[Predator555 0]

http://www.savefile.com/files/1580293

there is the file, im still working on the video, youtube is pretty slow today but the file should be sufficient for now

[KeyCat 131]

Tested your mission but don't know really...

The squad leader detects you when he turns his head to the left and sees you in the corner of his eye so I don't think it's a bug per se.

That said, the AI seems to have very good peripheral vision and maybe it could be tweaked to have a little bit less FOV to feel more natural.

Just my 0.20 SEK!

/KC

[Predator555 0]

i know what you mean but really i wouldnt expect to be spotted from that posistion and the fact remains that a 1 man patrol will never spot you yet the 3 man patrol always will, that doesnt sit well with me, it should be more random but yes perhaps describing it as a bug is possibly too strong

[jakerod 254]

i know what you mean but really i wouldnt expect to be spotted from that posistion and the fact remains that a 1 man patrol will never spot you yet the 3 man patrol always will, that doesnt sit well with me, it should be more random but yes perhaps describing it as a bug is possibly too strong

Well 1 head can look in only one direction at a time where as 3 can look in 3. Therefore the chances of them catching you is quite a bit increased. I would say that unless the patrol is not paying any attention you should be detected.

[walker 0]

Hi all

Not a bug just realistic show of human peripheral vision during normal human movement. Humans have a forward-facing 180 degree FOV.

BUT!

While walking humans look downward at about a 30 degree from the horizontal to make sure they do not trip on things.  By doing this Humans shift some of their peripheral vision to their rear. The angle of vision to the rear is thus about 270 degrees.

People also subconsciously scan left and right to increase their rear peripheral view. Merely turning your head 45 degrees to either side gives full 360 degree cover.

This is normal for humans and is almost certainly an adaptive response to our main predator threat's prefered attack angle. Big cats such as leopards like to attack prey from behind out of the bushes, crouching tigers and all that, most other predators including humans like that angle.

That said this is peripheral vision so it is tuned to movement. Though you will be aware of any blocks of colour and objects; you just cannot identify them but if you have identified them already you will label them in your spacial awareness which fools your brain into "seeing" the whole thing. Depending on your level of awareness you may turn to identify objects which your brain has not labeled.

How to achieve stealth in the test mission

Oh and by the way as to your stealth problem, just hide stood up in the little alcove to the right of the gate. If you face so that you rifle is paralel to the door you can get right  in it, so there is only a crack to see through; then lean to get view as the enemy walk past. Works every time.

Also having looked at the mission with a second squad member placed far away with both squad members set to "none" in the special drop down for the unit; so they are not in formation. I observed the column marching using command view and they turn to face and walk a few steps at 45 degrees off the line of march every so often.

Maybe this thread should be renamed "NOT an AI bug"

Kind Regards walker

[KeyCat 131]

That said this is peripheral vision so it is tuned to movement.

That was what I was thinking. For obvious reasons I don't know exactly how BIS are doing it but my guess is that the code only checks if a AI unit have LOS to target within a given FOV.

Of course a more natural behaviour would be to "filter" out not moving targets at the edges of the FOV. I'm sure BIS knows all about this but decided current way a good compromise, adding more calculations to each AI would maybe make the game run to slow etc. on todays hardware?

I guess that besides time/money developing sims/games is a huge balancing act between available HW resources and wanted features.

/KC

[rowdied 44]

Downloaded and tried it and I don't see a problem with the AI either.

Plus, I depboed it and tried it in the editor and noticed a few things.

1. With increased skill, they will spot you quicker and decreasing it they tend to spot you later. ( when the third man walks past )

2. Your positioning is not behind or to the side of the door but standing in the open doorway.

3. When you position yourself closer and farther back to the wall, the AI walks past and you can then ambush them. They do not see you when they can the horizon.

4. I've watched the AI and noticed that when they turn their heads to look, they will spot you out of the corner of their field of vision and turn to face the percieved threat and will only open fire when they have confirmed you are a threat. Hence the slight delay in raising and shooting their guns.

5. I placed myself in the bushes by the fence farther down the road and observed different behaviour. They all walked past but then spun around and searched the area for a percieved threat and proceeded to fired in my direction with guns and grenades. They only did this when I emerged from the bushes to get closer to the fence so that I could see them better. Again, they saw me from their peripheral vision and would 've kept moving if I didn't emerge from the bushes.

I see no AI bug problem but a problem of concealment here. If this was real life, I would be hugging the wall and not standing in the doorway, emerging only to kill them once they passed.

[Predator555 0]

you could all be right but to me it does not seem realistic.  just for the record i always hug the wall and do not stand in the open doorway but there are 2 issues.  Firstly i would not expect the enemies peripheral vision to be able to spot me considering both the distance and how much his head has turned.  Secondly even if i was within the range of his peripheral vision then that is no guarantee that i will be spotted, i think the distance is great and i am at best on the absolute edge of his sights.  when we consider that he is in a "safe" behaviour mode it is even more unlikely that he would see me, sure it may be possible, its also possible lightning could strike me down but not on a consistent basis.  The AI spots me almost every time

i have attached a video of me trying the test mission, i hope you take the time to view it and compare it to your own efforts and tell me if i am being unrealistic in my expectations of not being spotted.  maybe im nitpicking or just plain wrong, what do you all think?

here is the file, AI is in "safe" behaviour and difficulty is "normal"

http://www.savefile.com/files/1586885

EDIT - just rereading some of the posts (especialy in response to Walker) and im not looking for a way to beat the game by looking through cracks in walls and the like, i simply want to be able to stand in a realistic ambush position and have a good chance even if not 100%  of remaining unseen

[walker 0]

EDIT - just rereading some of the posts (especialy in response to Walker) and im not looking for a way to beat the game by looking through cracks in walls and the like, i simply want to be able to stand in a realistic ambush position and have a good chance even if not 100%  of remaining unseen

Hi Predator555

I am Refering to the open gate you are looking through at the start of your example mission. When you move into the alcove to the right of the gate in your example mission you can still see through the open gate but only smaller area I call a crack is visible. It is still the space between the two bits of the gate, it is just you see less of it because you are in the alcove pressed up against the wood.

AND!

You can increase the width of crack you see by leaning left. Leaning still keeps stealth high as more of your body is in cover but I tend to wait until they are about 5 yards past before leaning right out. Also this is hard cover and not just the concealment of abush so when you commence the fire fight you have the added benefit of sheild defilade.

By taking this position you are out of the 180 degree forward facing FOV and the 270 degree downward facing FOV and in the 90 degree wedge that can only be seen in peripheral vision where you 45 degree sweep. The predators prefered angle of attack.

For the record I just ambushed the column 3 times and have done it several times when I first posted as well. I did not get caught once. 100% effective ambush everytime. You wait until they are about 25 yards away and all in a straight column one behind the other; then take them all with 3 single shots or 3 bursts. If I had a silenced pistol I would normaly use that, straight to the Medulla oblongata tap tap tap. I think I will edit the mission and give it a go.

I will do you a video as an example.

Kind Regard walker

[Predator555 0]

ok, but you seem to be under the impression that i am looking for a solution to this mission, i am not.

i simply wish to know if it was fair to expect to go unoticed when in the position i took in the video i posted. to me if i did that in real life i would not expect to be spotted but am i wrong? or is it a fault of the game?

[walker 0]

Hi Predator555

You get spotted by the lead AI. The one you are tracking at moment 1.37 in the video.

It is a perfectly reasonable spot by the AI you are in the 180 degree Forward Facing FOV.

If you look at the lead soldiers head, he turns using the natural subconscious 45 degree sweep human beings are programmed by evolution to do. He spotted you fair and square.

In fact your video is an excellent example of how good a simulation of AI FOV and peripheral vision and movement vision ArmA has. I think it should be used as an example of how good the Neural Net based ArmA AI is when compared to other scripted AI.

As I said I think you should ask the modies to change the title to: "Not an AI bug."

You were however making several stealth errors that contributed to the spot.

You were after the shot too early. Buck fever hunters would call it.

You were too far out and for too long busting your concealment and you were moving far too much.

It is just maths x amount of body out of concealment multiplied by y amount of time adding z amount of movement equals that much more of certainty you will be spotted.

You were targeting the column leader rather than the tail of the column when attacking from the rear.

If you follow anyone one in the sight it is the second from last in a column where you are to the rear and you are only using him for a shot reference for height. Your 1st target is the rearmost then you shift target forward. This reduces the chance of a spot and yes ArmA models this.

All predators must learn stealth which is comprised of concealment/cover, patience and stillness; whether you are a kitten learning to hunt birds in the garden or a sniper.

First things first.

Bring up your watch default key [o] [o]

PLAN your shot. You know you are going to ambush them. You know you want them in a compact group. You know they are in column. You know the road curves and dips to the right. At that curve they will all line up as the front man turns to the right and the others in column follow; the column will will fall into the perfect predators angle.

Average human walking speed is around 1 metre per second. You have identified your first reference point in you video it is the wooden road edge marker at the road bend at 30m from you. The Targets need to be 25 to 30 m from you position for that prime shot. The Targets begin about 80m from your position. 80 plus 30 gives 110 seconds you could add in about 5 for side paces and dawdling or looking around.

The time you need to shoot is about 1 minute 50 seconds why look before 1 minute 40 seconds? The target will not reach the shoot mark before then. And all that time your head and your sticky out helmet and gun with their unnatural straight hard lines will be bobbing and weaving in the targets peripheral vision.

You know where your target will appear so just be patient. Do no bob you head, You are using TrackIR which is excellent but you need stillness not awareness. Shake out any tightness in the neck while right up in the corner but at 1 minute become still. Take slow breaths. There are various techniques hunters and people in other activities use to attune themselves at moments like this. Learn one of them. Basically do what a cat does in moments like this become still.

Absorb your self in anticipation of the moment.

Rehearse the shot in your head several times. See the targets all lined up in a column. See your red dot on the top of the neck of the back target. See your self taking a breath and holding it see your self taking that first shot and the target going down. See your self quickly lowering the red dot to the lower (the road dips remember) 2nd target and taking him in turn 1 shot maybe two. 3rd Target lower still and aware and moving possibly even hitting the dirt, switch to burst if need be to take the last target.

The above process is called visualisation.

At 1 minute 40 seconds lean out.

Do not bring up your sight too early obscuring your field of view and situational awareness bring the sight up at 1 minute 45 to 50 seconds. Use your rifle against the edge of the wall to produce a smooth turn on to the second soldier in the line. The base of his helmet is your second reference point. Now wait hold that height and wait for the rear most target to walk into your red dot.

Do the shots as the targets line up; exactly as you rehearsed them in your head.

As I said I am managing this 100% every time.

Kind Regards walker

[dwringer 45]

Haha, well it looks like this topic has been discussed to death already, but I just wanted to point out that the enemy patrol is FAR more likely to see you since they are trained soldiers - trained to look for anything out of the ordinary, and certainly any enemy troops lurking around.

This, as with all other criticisms of ArmA's AI that I've had myself, falls under the general complaint, "ArmA's AI isn't as smart as actual humans". Which it can't be, considering the amount of CPU that would eat

[KeyCat 131]

maybe im nitpicking or just plain wrong, what do you all think?

It may be close to "nitpicking" but I see nothing wrong to have a proper discussion on how to improve AI (and other things) in ArmA.

I quote Q and The-F from their PROPER AI Test Framework readme which I also agree with. I think it's kinda unrealistic to expect that BIS would have resources to tweak every value to perfection...

Quote[/b] ]

The background for this project is that like in OFP in our view the AI in Armed Assault has still much room for improvement via config tweaks.

And who knows, while tweaking and testing different things some may unveal a real bug/issue that may be corrected in a future patch or ArmA 2.

I currently don't see this as a bug per se but after playing around with your mission I think the AI's current peripheral vision could be toned down a notch or two since IMO they are a bit to good to detect static targets in the corner of their eyes.

I'm looking thru the config files and can't find any value to to adjust the AI's FOV ( am I missing it? ) but there is a value that sets their min/max headturn in degrees (which seems to be 120 degrees for soldiers).

<table border="0" align="center" width="95%" cellpadding="0" cellspacing="0"><tr><td>Code Sample </td></tr><tr><td id="CODE">

 minHeadTurnAI = -60;

 maxHeadTurnAI = 60;

I will play around with thoose values and see if things improves or not.

A thing I noticed while testing is that sometimes the AI's head movement becomes kinda spasmatic when they scans left/right, no big deal but looks kinda funny and just telling what I observed...

/KC

[Predator555 0]

maybe im nitpicking or just plain wrong, what do you all think?

It may be close to "nitpicking" but I see nothing wrong to have a proper discussion on how to improve AI (and other things) in ArmA.

I quote Q and The-F from their PROPER AI Test Framework readme which I also agree with. I think it's kinda unrealistic to expect that BIS would have resources to tweak every value to perfection...

Quote[/b] ]

The background for this project is that like in OFP in our view the AI in Armed Assault has still much room for improvement via config tweaks.

And who knows, while tweaking and testing different things some may unveal a real bug/issue that may be corrected in a future patch or ArmA 2.

I currently don't see this as a bug per se but after playing around with your mission I think the AI's current peripheral vision could be toned down a notch or two since IMO they are a bit to good to detect static targets in the corner of their eyes.

I'm looking thru the config files and can't find any value to to adjust the AI's FOV ( am I missing it? ) but there is a value that sets their min/max headturn in degrees (which seems to be 120 degrees for soldiers).

<table border="0" align="center" width="95%" cellpadding="0" cellspacing="0"><tr><td>Code Sample </td></tr><tr><td id="CODE">

 minHeadTurnAI = -60;

 maxHeadTurnAI = 60;

I will play around with thoose values and see if things improves or not.

A thing I noticed while testing is that sometimes the AI's head movement becomes kinda spasmatic when they scans left/right, no big deal but looks kinda funny and just telling what I observed...

/KC

thanks, i noticed the spasmatic ai head movement as well, its a little strange.

i agree with you, no matter what others have said i just find that at the edges of the AI sight are to wide and as clear for them as something dead in front of them

as for the issue of peripheral vision, well i highly doubt this was modelled for the AI, they will just have a sector of vision that they can see, simple as that but if someone can point out that BIS actually did model it then im happy to be proved wrong

perhaps the issue is exactly that, the AI has normal vision where there should be a blurred and unreliable peripheral vision at the extreme edges of their sights.

[mrcash2009 0]

LOL ... I can conclude that the "spaz head" does exist here too.

I completely ignore it and kind of find it funny, but then again its not meant to be that way

[walker 0]

Hi all

I think KeyCat has the correct approach on any analysis of ArmA AI. The scientific one.

Factors in the spot equation:

Hiding subject side:

Amount of subject that is visible

Amount of time subject is visible

Amount of movement of subject

In this case use a camera placed at the spot location to get these values in MP or just the first two in command mode in SP then asses the movement in the video Predator555 provided for the last.

Add in silhouette value by comparison of values for High Dynamic Range histograms for subject and their background.

Add in Flash from Sun angle. Consider that if you move the sight between the sun angle and the target you are certain to hit the flash point. Another reason not to move and minimize sight tracking.

Target Side:

AI static FOV asume180 same as humans

Movement head turn 45 to -45

Total movement plus static FOV = 270

AI allowed head Turn 60 to -60 (This is what I can turn)

Total head turn plus static FOV = 360 what humans are blessed with by evolution in response to our predators.

I think this is what caught you.

How often a soldier would do a full head turn as part of Patrol SOP. There is probably a value for this in the AI config.

In NATO you would do a full body turn every so often too.

I have seen AI do this in Aware mode.

What are the awareness abilities of the target? These are values set in your ArmA config and by you in the editor.

To me the values seem correct but we are talking glass half full versus glass half empty here. I think the values are valid ones but if someone wishes to pay me to do the research I can analyse and verify it for them whether it is enough to meet an accreditation procedure depends on what those criterion are.

Kind Regards walker

[KeyCat 131]

To try to gather some more data/facts regarding this topic I made a simple addon that disable the default 60 degree left/right headturn on east AI.

I haven't had much time to test it more than verifying that the addon works. Use this (test) addon with Predator555's test mission above and post your observations.

My limited tests shows that also without being able to turn their heads +60 to -60 degrees they are still extremly good to detect your presence by using their peripheral vision (remember that we are standing completely still hugging a wall in camofluage clothes) IMO in an un-natural way.

Anyway, test and post your thoughts in this thread...

http://keycat.no-ip.com/files/AI_No_HeadTurn.zip

Note: This addon is for test purpose only and only modifies east AI soldiers.

Last night I was thinking about the built-in "peripheral vision aid" we as players gets while playing (the white/green/red light orbs at the edges of the screen). I'm just guessing here but maybe also the AI somehow benefit's from this "extra sense" and thats what makes them feel a bit superhuman when it comes to their peripheral vision?

/KC

[walker 0]

Hi all

As prommissed a video showing how to do the ambush and not get spoted by moving into the alcove.

http://www.thechainofcommand.net/downloads/videos/Not_an

Also shows how the AI looks to the left and spots you when you hide nearer the gate.

Kind Regards walker

[KeyCat 131]

Nice video walker and I agree that if staying well hidden behind solid objects like walls etc. the AI will not detect you. Thats why I personally was unable to reproduce this issue when it first came up in the 1.12 Beta thread and before Predator555 posted his test mission.

What I'm trying to find "evidence" for is if the AI may have a tad to good peripheral vision or not when it comes to static targets. In this particular test where you are standing still and only partly hidden behind the wall and no silloutte against sky it feels like they do.

In previous post you mention object movement, silloutte etc. do you (or anyone else) have any info/pointers to how much the AI code actually take into consideration when calculating this stuff or are you just guessing/assuming?

As said before, this may be nitpicking but any fix/tweak/mod/whatever that makes the AI behave more like we do (and that includes all our limitations) is IMHO a step in the right direction and AI in general is a quite interesting topic as well.

/KC

[walker 0]

Hi KeyCat

Well you have set the question. Here is the proposed experiment.

Baseline

I think the first thing is to establish a baseline.

What can a human player see from the same position?

Procedure

So the easiest way to do that is to put two units in as playable same side then using Team switch set up the experiment.

Limitations

There may be a problem that requires FOV to be increased which is that normal FOV for ArmA is reduced in ArmA as in all games due to to limits of computer monitors a Triple head display would be the best option BUT we can just alter ArmA player FOV to match a 180 Degree Forward Facing FOV using fovTop and fovLeft so you can see two objects at 90 degrees to either side of you. Of course Humans actually look down to get a rear view and move their eyes and sweep as part of normal human movement.

[Edit] OK Just tested it and this one is, bug bust.

Even without matching Human FOV! Just using vanilla ArmA settings of

fovTop=0.750000;

fovLeft=1.000000;

If you angle of march is the same as the AI in Predator555's first video 260 to 280 Degrees.

Zoom out Default Key (Double click number pad minus [-]) to get maximum available view in ArmA. Which is still only about 75% of Human FOV.

The soldier in gate is plainly visible. More than Half their body in fact. Their Camoflage helps but any head movement draws you straight to them.

I can Definitively say this bug does not exist. Maybe the title Should be Changed to Not an AI Bug.

'E's not pinin'! 'E's passed on! This bug is no more! He has ceased to be! 'E's expired and gone to meet 'is maker! 'E's a stiff! Bereft of life, 'e rests in peace! If you hadn't nailed 'im to the forum 'e'd be pushing up the daisies! 'Is metabolic processes are now 'istory! 'E's off the algorithm! 'E's kicked the bucket, 'e's shuffled off 'is mortal coil, run down the curtain and joined the bleedin' choir invisibile!!

THIS IS AN EX-bug!!

[/Edit]

Kind Regards walker

[Predator555 0]

appreciate the first video

anyway can you post a second video of your proof that it is not a bug, id be happy to be proved wrong

EDIT - (btw i hope you have my position correct, use the video i posted, the position you start in the test mission is not a position i would expect to go unnoticed from, in the video i clearly move to a slightly better position, more closely tucked into the wall and peeking)

link is still the same - http://www.savefile.com/files/1586885

[walker 0]

Hi all

As requested a video showing how the ambush position is compromised.

http://www.thechainofcommand.net/downloads/videos/Not_an

That said I have described the process of the experiment and it is repeatable so that people can confirm or dispute my findings. Remember videos can be faked but experiments are repeatable and testable.

I would also like to say the process in this thread has been excellent in the way all parties have reviewed the evidence and presented evidence. This the proper peer reviewed way bug tracking should be done. Even those that are only an apparent bug rather than a real bug.

I congratulate Predator555 for setting a proper tone in this thread and presenting a proper scientific case with video and example mission rather than just a rant as some in other threads have.

I hope more people will follow his excellent example.

Kind Regards walker