AI Discussion (dev branch)

[dragon01 902]

I think that the real problem is that AI doesn't have a concept of "disengage". If they see the enemy, they've got to fight, period. They practically stop following other orders, too, which is annoying. It'd be a great leap for ArmA to introduce a "turn tail and run" combat mode that could be ordered by the player. The "stealth" mode could place more emphasis on actively avoiding combat, as well.

 

What ArmA needs is more reliable ways of ordering the AIs around. This includes disengaging, exact waypoint following (for example, if you send an AI somewhere, he will engage, but won't move from that spot unless ordered to stop holding it), fine stance adjustment (a great feature that the AI seems to have no concept of) and better behavior in confined spaces. I think the latter would be best achieved by detecting when the AI is entering a "confined space" and switching to a completely new "behavior model" more suited to urban combat. Old Rainbow Six games and the first three Ghost Recons (the latter were not too shabby, either) are a great example of AI behavior done right.

[x3kj 1247]

 

I think the latter would be best achieved by detecting when the AI is entering a "confined space" and switching to a completely new "behavior model" more suited to urban combat. Old Rainbow Six games and the first three Ghost Recons (the latter were not too shabby, either) are a great example of AI behavior done right.

that a different behaviour in confined space compared to free environment is required is without question. But what exactly makes a good behaviour? Remember that not all AI behaviour is the result of the logic. In many games the developers map out all the points of interest (low cover, high cover, doors, ...) for the AI in their pathing that is defined in the map. The code then just lets the AI go from one predefined location to another. This is often very apparent in such games. In the majority of A3 buildings, there are no cover objects inside houses other than doors and walls.

 

The only thing i could think of would be to define a set of PathLOD positions as a room, so the AI knows what positions belong to one single room. Within one room set the doorways (or "chokepoints") that lead to other rooms could also be marked, so that the AI knows in which direction it needs to cover (rather than aiming in a random direction, sometimes facing the walls). Window positions inside a room could also be marked, so the AI knows what position to take if they discovered enemies outside the building.

[fn_Quiksilver 1636]

need that 'forget' command counterpart to 'reveal', or allow 'reveal' to reduce knowsabout

[froggyluv 2136]

I think that the real problem is that AI doesn't have a concept of "disengage". If they see the enemy, they've got to fight, period.

 

 

 Although very much agree with the"indoor fighting" part of your post, its this above that would really start allowing a new dimension of combat. As it is now, enemy AI sees you, it fires. There is no feeling of being stalked, watched, hunted, or anything remotely offering of that realm and that is a huge shame as it would literally open up the game on  not only a strategy game level, but add intrigue to the more 'story narrative prose' among us.

 

 To me the biggest disappointment is that the past release tells us not only do they not care about these advances, but that they are going in the diametrically opposed direction of controlled storyline and MP. In short, give it up lads the above will NEVER happen.

[dragon01 902]

Except that BIS devs are listening to us. Really. I remember when I suggested to have the "Loiter Altitude" available as an option in Eden. It literally got implemented during the next update. :) So don't say that it'll never happen. If we yell about it loudly enough, it will. Convince BIS that Apex was a mistake, that we want the AI improved, not ignored. Show that there's a large demand for it and it will come.

that a different behaviour in confined space compared to free environment is required is without question. But what exactly makes a good behaviour? Remember that not all AI behaviour is the result of the logic. In many games the developers map out all the points of interest (low cover, high cover, doors, ...) for the AI in their pathing that is defined in the map. The code then just lets the AI go from one predefined location to another. This is often very apparent in such games. In the majority of A3 buildings, there are no cover objects inside houses other than doors and walls.

 

The only thing i could think of would be to define a set of PathLOD positions as a room, so the AI knows what positions belong to one single room. Within one room set the doorways (or "chokepoints") that lead to other rooms could also be marked, so that the AI knows in which direction it needs to cover (rather than aiming in a random direction, sometimes facing the walls). Window positions inside a room could also be marked, so the AI knows what position to take if they discovered enemies outside the building.

Expanding on PathLOD is a very good idea. I think that in general, when possible, objects (not only buildings) should have pre-defined positions to make it easier for AI to find its way around them. For example, a low wall could have a number of cover points, each specifying a stance for shooting, stance for hiding, direction of cover and its type. The first Crysis, IIRC, did something like that and it was praised for very good AI. I think that lessons should be taken from the few other games which have done a similar thing to ArmA, but on a smaller scale.

[x3kj 1247]

 

For example, a low wall could have a number of cover points, each specifying a stance for shooting, stance for hiding, direction of cover and its type. The first Crysis, IIRC, did something like that and it was praised for very good AI. I think that lessons should be taken from the few other games which have done a similar thing to ArmA, but on a smaller scale.

This will be very problematic. Buildings are buildings, and there is generally nothing inside them. (whenever you do place something inside them, it becomes a big problem for path finding, or AI might just phase through those objects). When you expand this to exterior objects as well, like walls you would have to double check everything:  what's the best pre-defined point for cover (which would require also to code in from which direction the cover is suitable), and also to check if that point is even valid? Because there might be another object that blocks the position. I'm certainly not against that idea, but i think it would be more problematic and timeconsuming than doing only the interior of buildings (and all that counts as buildings, like bridges etc).

 

Improving the pathway LOD is just providing more useable information for AI code obviously. It still needs AI code improvement to be worth anything...

[dragon01 902]

I think that the problem you're talking about would be easily solved by adding an obstruction check to a position. Indeed, building interiors could use that, too, because they'd usually have something placed inside them if you're going to make use of them. Of course, that'd require accounting for it in patchfinding, too, but if you're going to have mission-critical buildings, you can't have them devoid of literally anything.

[seba1976 98]

I know I'm taking a wild guess here, but I think more is not being done in the AI department, because the devs know they don't have any spare computational power left, and therefore can't implement anything that put any more stress on the engine. When they finally change the engine to something multithread, I'm hoping to see big improvements in this regard, and maybe a whole revamp of the AI system.

[Vasily.B 529]

I know I'm taking a wild guess here, but I think more is not being done in the AI department, because the devs know they don't have any spare computational power left, and therefore can't implement anything that put any more stress on the engine. When they finally change the engine to something multithread, I'm hoping to see big improvements in this regard, and maybe a whole revamp of the AI system.

In arma 4 :)

[44000zom 13]

Anyone else find that while the AI are in stealth mode and over grass they insist on crouching instead of going prone? This is great for other the behaviors but it's frustrating when you go prone and the rest of your squad decides that they want to be spotted by remaining crouched! In stealth mode you would think that remaining hidden and therefore prone in grass would be the priority. Setting them to go prone manually every time is slightly annoying and I'm sure it used to work fine a couple of months ago?  

[inlesco 233]

Has anyone ever wondered why... AI never reports friendly fire on vehicles they're in or driving, unless you hit some unit?

 

You can easily unload an entire MG mag into a vehicle and nobody says a word :D. A rebel can dream.

 

So yeah - I'd love to see AI vehicle FF implemented sometime.

[Guest]

 

Imo, the elements of Ai that can pull off being able to hunt and stalk is tied to two things, refined/detailed non-group oriented style Ai behavior and refined mission making that includes more detailed and refined mission areas.

 

This is not so aimed at units not firing if told to hunt/stalk their target, but rather units in general that have the ability to advance upon known enemy positions 

that are 'not' in sight of their targets - of which such units in volume can be quite a few often, specifically in major mission areas that have enough 'stuff' around 

to break things up.

 

AI Modify actually does pull of quite a bit of this, but it still requires an effort put into a particular mission. Like with the few other games I have seen this type 

of ai behavior, there is no fully 'autonomous' Ai that can do everything itself, it requires a good Ai system that provides a good amount of control over the Ai's more detailed actions for mission making along with good effort put into hand working mission areas.

 

The reason I say 'non-group' Ai behavior, is because units are individually doing a variety of different things including picking up known enemies via Los and Fov of other nearby allies that know of enemies, I couldn't even imagine trying to tie all of that into orchestrated group behavior.

 

When the automatic "'all units in groups' instantly know about an enemy when one unit in the group knows about an enemy" is *removed*, and replaced with all units only being able to pick up information from other allied units depending on Los, Fov, and distance to each other, this immediately opens the door to really be able to start creating some mission scenarios that can unfold quite well.

 

Supposing there are 50 units in a medium sized base. A player group with suppressed weapons quietly infiltrates the entrance, but then one of the units further in sees one of the players before being killed and there is another nearby unit further in that sees that unit  and picks up the players position  - and then hence a chain reaction occurs where say 3 other units nearby the unit that sees the unit get killed which have Los to the unit also become alerted and gain the last known enemy position - Now there are 4 units in the base that have been alerted, have a last known enemy position, and hence they begin to move towards the last known enemy position - this movement likely involves different paths, with the units keeping distance between themselves - they slowly weave between buildings and walls, using corners at times to check for clearance before advancing further. If other units still in safe mode at other places end up seeing these guys moving to an enemy position, they will switch to combat mode and have the ability to pick up those units last known enemy position and advance to it as well. And when I say 'last enemy position', I mean stored last known enemy positions for each unit which could be realistically quite outdated as time goes on - this is the way it should be, NOT just simply transmitting the players current position in an undesirable generic fashion. And of course, gunfire sound has its own direct effect, where an enemy unit firing a loud weapon will alert a very large area around him, and the same for the players, except that suppressed weapons only have a 30 meter range. Units also have control over how far they can advance on an enemy position, along with other things as well - which means of course that there is a level of unexpected behavior in this area.

 

The above result, in a well done hand worked mission area, can certainly provide the element of being hunted down be the Ai often, and helps to keep players moving, as getting pinned down can get nasty. A hand worked mission that spreads out the overall enemy Ai volume over a given area often can help to lead to scenarios that unfold where the enemy can end up advancing on the players from various directions as well, provided of course there is plenty of 'stuff' around to break things up, primarily meaning buildings and walls, but of course other things as well.

 

This can also be accomplished in more outdoor style areas, but again, these areas need to have 'stuff' to break things up, trees, boulders, varied terrain over short distances would be nice although this type of hand worked terrain I've never seen in the game, and other clutter types including building ruins, etc. One thing about outdoor areas that I think is important is to keep vegetation (bushes,etc) to as much of a minimum as possible, and try to stick with more solid objects that cannot be seen through nor fired through - this helps to significantly avoid forest combat that can become quite negative and annoying due to excessive foliage that one has to keep straining their eyes to see through in order to see the enemy before they see you, over time this can start feeling more like a job than a game.

 

Anyhow, my point is, Imo to a degree it can be done in Arma 3 - but it Imo cannot be done strictly autonomously, probably should be a separate Ai system used only in smaller scale missions because of the Cpu usage from 3d line drawing and other script commands running, and lastly still requires a good effort in the making of missions using such an Ai system in order for the Ai to provide a good quality effect.

 

Lastly, this type of refined Ai behavior really both detracts from the open war-field, group oriented combat that Arma 3 does a good job of providing, which is why Imo these should be two separate systems, for separate mission types. The added bonus to a separate Ai system is also that mission makers and modders can choose to use it or not, which means any hard coded stuff in the system won't mess with mods and missions if the authors don't want to use it.

 

I personally think that Arma 3 is the last remaining combat style shooter that has still kept to a lot of great basics : Dynamic Ai, Free leaning (no sticking your 

avatar to predefined corners to shoot), free open style mission environments, many vehicles that can be used,and a ton of other stuff - and of course the huge volume of great modded/mission content out there. This is really why I personally see this as really the only game left that is worth putting the effort into a separate 'refined' Ai combat system for more personalized and what would likely often be special ops style missions. Even if Hidden And Dangerous 3 ever happened, it probably would have little if any modding options anyhow, or else it would probably be my better choice if it did only because I am still impressed with what was done with HD2 overall (even though the AI does some ridiculous stuff sometimes).

 

 

[Bitesrad 481]

Is there any way to reduce AI accuracy even further? I'm trying to make make some setups in the editor with lowly trained militants, but even with minimal skill levels they're still a bit to accurate for my liking.

[joostsidy 685]

I think that the real problem is that AI doesn't have a concept of "disengage". If they see the enemy, they've got to fight, period. They practically stop following other orders, too, which is annoying. It'd be a great leap for ArmA to introduce a "turn tail and run" combat mode that could be ordered by the player. The "stealth" mode could place more emphasis on actively avoiding combat, as well.

A while ago disable 'AUTOCOMBAT' was made possible, which prevents autoswitching to combat mode. It should help with turn tail and run behaviour.

 

I do not agree with the statement that AI never stalks, hunts etc. If you engage one or two enemy squads by yourself with a distance of a couple of 100 meters and lots of cover, you can get into trouble because the enemy will supress you with their LMG's, while sending individual soldiers to flank you or even attack you from behind. I've got myself into trouble like this multiple times, underestimating the ai and loosing track of the enemy soldiers as I need to take cover a few times and their soldiers disperse seeing them again only when they appear nearby.

[fabrizio_t 58]

AI definetely improved across time.

 

2 years ago you needed mods in order to get suppressive fire / suppression, morale penalties impacting behaviour and accuracy, consistent use of frag / smoke grenades.

Now vanilla AI provides these features.

Moreover targets detection has been revised and now works better, as well as pathfinding.

Sure, there are still issues, such as problems with convoy driving, firing accuracy glitches, occasional hiccups, problems with rearming and healing, etc.

 

Yes steps were done in the right direction.

[fabrizio_t 58]

I think that the real problem is that AI doesn't have a concept of "disengage". If they see the enemy, they've got to fight, period. They practically stop following other orders, too, which is annoying. It'd be a great leap for ArmA to introduce a "turn tail and run" combat mode that could be ordered by the player. The "stealth" mode could place more emphasis on actively avoiding combat, as well.

 

What ArmA needs is more reliable ways of ordering the AIs around. This includes disengaging, exact waypoint following (for example, if you send an AI somewhere, he will engage, but won't move from that spot unless ordered to stop holding it), fine stance adjustment (a great feature that the AI seems to have no concept of) and better behavior in confined spaces. I think the latter would be best achieved by detecting when the AI is entering a "confined space" and switching to a completely new "behavior model" more suited to urban combat. Old Rainbow Six games and the first three Ghost Recons (the latter were not too shabby, either) are a great example of AI behavior done right.

 

I've ranted so many times about AI non being able to properly withdraw / flee over years :)

It's amusing AI units still flee in formation.

 

I agree that "stealth" mode should make AI more "stealthy", meaning avoid sprinting and idle standing / crouching when enemy is in sight.

 

Not sure at all about the need of further / tight orders, i think micro-managing is already too much.

[dragon01 902]

I think that first of all, it should become possible to micromanage them so that they don't get themselves killed. IMO, it's crucial that in a tight spot, they do exactly what you want them to do. This shouldn't be mandatory, but no matter how good the AI was, you'd always have a situation sooner or later when it's just not good enough.

[fn_Quiksilver 1636]

After watching players use recruitable AI on my server:

 

 

AI commanding menu should include some disableAI toggles.

 

Specifically:

 

unit disableAI 'PATH';  // for garrisoning units, commandStop works, but unit can be 'spooked' out of this state

 

and the important one

unit disableAI 'AUTOCOMBAT';

 

^ That one really needs to be on the commanding menu.

 

In combat/firefight I want the AI to work through its decision tree normally. When I want them to move/follow regardless of context, I want to disable the autocombat temporarily.

[fn_Quiksilver 1636]

I want ability to have units disregard formation direction without disabling AI components.

 

Reason:

 

Currently when you garrison units of the same group into buildings, they all rotate according to formation direction. We can fiddle around with doWatch, but I want to simply set the direction of the unit via script and not have them twist back to formation direction.

 

 

 

Also what does this command do? https://community.bistudio.com/wiki/resetSubgroupDirection

[inlesco 233]

I want ability to have units disregard formation direction without disabling AI components.

 

Reason:

 

Currently when you garrison units of the same group into buildings, they all rotate according to formation direction. We can fiddle around with doWatch, but I want to simply set the direction of the unit via script and not have them twist back to formation direction.

 

 

 

Also what does this command do? https://community.bistudio.com/wiki/resetSubgroupDirection

I fully support this and all your ideas on enhancing AI functionality!

 

BTW, is there any way to force AI ignores obstacles in its way when on foot? AFAIR, there's a command to ignore obstacles (in other words, execute less code) when driving, but there's no variant for on-foot behavior.

[mickeymen 324]

Hi BIS!

 

Still, ARMA 3 needs a way to get subordinates soldiers to move quickly to the specified point, because any automatic-switching subordinates AI in danger mode, blocks the ability of normal movement!

 

If the players AI-squad, during movement switched to danger mode (note! this can be, even if the threat is not big), then this AI-squad always stops movement. He stops or falls to the ground and does not move in the set point, notwithstanding any commands of player. Any squad! Under players command or AI-command, will not continue normal movement, up until the commander will not say "Clear".(i.e. exit from danger mode). 

Only after this point, this AI-squad, will be able to continue the movement to the destination point. This feature of AI really slows down the any movement to the any destination point and can even be fatal for the whole squad. Sometimes, period of inability of the AI to move, takes much time even if squad no has a real threat. Several and more minutes, the AI-squad not able continue its movement.

 

In real combat, there may be situations, when soldiers should only run and don't get stuck down in the combat!  But for unknown reasons to me, this is ignored in Arma. For example, this can be a - breakthrough towards to the priority target or running to the helicopter in order to save their lives and leave the dangerous region, or - break out of the encirclement of enemies, or - salvation from a enemy mortar attack, as we can see in the singleplayer mission, when Kerry should be to run through mortar fire and etc, etc. 

But, unfortunately today (v1.62) in ArmA3 does not exist any type of movement, in which the AI will not be stuck up in the battle and will not stop before the specified target point! Fully absent any type of movement in which all bots will move to the specified point with the highest priority!

 

I propose to add a new type waypoint, which will be called "BREAKTHROUGH".

In this case, the AI will have to trying to reach the set point as quickly as possible, of course risking more. 

Also this will allow to add this command in the actions menu to the player-commander could command to do it.

For example movement commands can be divided into 2 types: 

 

"MOVE TO HERE" - will standard movement as we see today (v1.62). 

"BREAKTHROUGH TO HERE" - will movement with highest priority.

 

This addition, could be a new tactical choice for every player as commander and make the AI units more flaxible and interesting.

 

BIS, I think this is not difficult task as can looks, in such moments, you need to turn off or almost disable FSM for AI and make the AI-movement as quickly as possible(sprint). Every command - "return to formation", will get back to their normal state. Or if the stamina for the sprint to an end, then AI must to return to the FSM back.  These are just some my suggestions. Please make a BREAKTHROUGH possible for AI.

Any ways, welcome...

[froggyluv 2136]

 I hear what your saying and weve been asking for the "Run At All Costs" command for a looong time. They did however recently give us the ability to disable AutoCombat and that does in effect what your saying to a point. For instance, was playing last nite when me and my AI squaddies were traipsing thru the woods stumbled upon a BMP which got alerted and started to rotate its turret. I disabled the AutoCombat with quick key press (a mod), put them in alert and told them where to run -which they did. Hopefully BI will one day add this themselves with even more advanced commanding but i wouldn't hold your breath

[fabrizio_t 58]

Currently when you garrison units of the same group into buildings, they all rotate according to formation direction. We can fiddle around with doWatch, but I want to simply set the direction of the unit via script and not have them twist back to formation direction.

DoStop each of them and set their direction.

[Vasily.B 529]

@mickeyman - I agree about breakthrugh option - but i know this may be resolved in other way. Ac you place waypoint in editor, you can choose behavior of AI, they can fire at will, respond with fire, or ignore fire - this could be applied/improved in behaviors.

 

For example i remember (not sure if its still present ) situation, when i ordered my man to hold fire, and move in some position (they were in combat mode) - what did they done? They was moving, crouching, aiming at enemy (without firing), keep moving further, crouching, moving etc. Their behavior was.... funny.

 

Why they was aiming at enemies, when i ordered them cease fire? More over, you cannot change their behavior, once they enter "combat mode" to "Aware" and lower.
Thats only my 2 cents.

[mickeymen 324]