gameplay-ai

[celest python]

[rich veldt]

Hello! I'm trying to study EQS and AI spatial awareness a bit more and figure out what I'm able to do.

I have a basic "find hiding spot" EQS which gets locations on pathfinding grid, does a line trace test to the AI's CurrentTarget (custom Context), and filters out any items that are visible to that actor, and then scores the non-visible items based on distance (preferring the closest to the querier).

It works well enough, the AI runs behind (ish) the nearest wall.
On some occasions, the AI might be totally occluded by the wall, and on some occasions I guess the winning item is JUST barely occluded from the CurrentTarget so the AI sticks out and is hardly "hiding".

My goal here is to get the AI around a corner or something and have them hug the corner, sort of like Condemned's enemies, where they are baaarely visible (and maybe then play a "leaning out" montage)

I have a few ideas on how I could get a more suitable position behind the wall, like having a custom trace test which doesn't trace from the item's position, but something like 20-30uu to the left and to the right as well to simulate the total width of the hiding character, and only returns items which have blocking hits on both/all trace tests.
Then have the grid items be generated more densely to get a more accurate item.

I'm just wondering about anyone's thoughts on how this could be achieved, or if I'm barking up the wrong tree here or missing something with EQS and behavior trees that would make this far easier than my custom test idea.

[ocean wren]

Personally, I dealt with this by generating cover points as a seperate process and then used EQS to select from the cover points with specific characteristics. So you can make a cover generator select cover with a specific height, with visibility (or non-visibility) to a location, with step-out wall cover etc. Then use EQS to choose from them. I also have a claim system on the point so that more than one AI doesn't go to the same spot.

I didn't feel like making EQS do it all was going to work

That's not to say it couldn't with enough finessing.. but I was mainly just doing what other games did

copied mostly from cryengine in this case 🙂

[rich veldt]

I've considered this method too, but if "cover" in this case is any corner in a hallway, pillar, etc the AI can put their back up against, what's a method for generating cover points? Perhaps a similar method to what I suggested but not at runtime and instead something generated in-editor? Generating points around the navmesh, running linetrace checks on those points to see if they're near a wall, then on the left and right to see if they are indeed near a corner?

Then I guess collecting all "winning" points and spawning the coverpoint actors with necessary characteristics such as a "claimed" bool etc like you mentioned

I certainly would prefer doing that I was just having a hard time conceptualizing how the generator would be coded. Finding convex corners is certainly more complicated that concave corners lol

[viral field]

Hey guys, does my ai need a real behaviour tree to get nav links working, or does a move to actor/location also work? My ai doesnt fall from platforms. Walk off ledges is activated

[misty wharf]

regular move to will use nav links just fine if they are set up correctly

A behavior tree is nothing but a way of organizing logic

[viral field]

hmmm, i wonder whats wrong with my nav mesh / nav links, its not working

[misty wharf]

Is it not moving through the links or is it getting stuck when it attempts to enter the link?

[viral field]

It looks like its moving to the first point, but then doesnt drop down. You can actually see that its supposed to fall, but it doesnt. So either i have to make a new colission capsule, because its a sphete right now, or something else is wrong which i cant figure out. I can send you pics as soon as im home

[misty wharf]

Yeah I'm sure someone can help you with that if you provide a bit more details :) If it's just for falling off a ledge, then it seems like as long as the positions of the start/end and the settings are correct it should work as long as your CMC if you use a character is set up to allow it to walk off ledges as I recall it has a setting that prevents it

[viral field]

What is cmc? And yes walking of ledges is activated

[misty wharf]

character movement component

[viral field]

ah ok thanks, so i will take some screenshots now

here some pictures

movement component is default ( its the default pawn blueprint )

*movement component settings are default

I noticed with the behaviour tree, it moves down like it would be a ramp, buts not supposed to be like that :p

[terse bear]

I'm using node memory in my service, does it lose the data when the branch evaluates or when the whole tree is re-run?

[viral field]

sorry i dont know what you are talking about :S

[celest python]

shouldnt be

[viral field]

im currently using ai moveto anyway, since the tree would somewhat work, but not as intended

[celest python]

check visual logger

https://docs.unrealengine.com/4.27/en-US/TestingAndOptimization/VisualLogger/

[terse bear]

Wouldn't running the tree from the beginning re-assign the node memory?

[viral field]

so the recast default says initpathfinding start point not on navmesh. the controller says trying to find path to "ai name" resulted in error

ok looks like its only happening when im jump. so basicly everything is blue, it also shows my position on the debugger

and the ai's path

as soon it reaches the navlink point, it get "trying to find path to "ainame" resulted in error

[viral field]

ok i guess i have found the problem, the event tick was causing the troubles. But now its just floating down, not realy falling :/

it somehow slows down if reaches the point

[rich veldt]

Nvm I did some googling and found a nice one by Deam. He says it could be better optimized, and I suppose that could be done by having it generate only in a small area and only when enemies that use cover are in that area. By the looks of it I could probably edit it to do that.
https://github.com/Deams51/CoverGenerator-UE4

Really clever use of the nav mesh.

[celest python]

there is also this: https://github.com/GlassBeaver/CoverSystem

Glass doesnt support it anymore though

but its still one of the most advanced ones ever released so far

[celest python]

not sure but unless you re-set the BT on AIController it should keep the data

i.e. change the braincomponent (what UBehaviorTreeComponent is) or asset

[elfin hinge]

Hi. What's the better way of controlling multiple (around 50) AIs in the arena? Each one have their own AIController or have a single controller that is controlling each and every one of them (even if their types are different, all of them use same class AIController).

[celest python]

if you need a manager for your AIs, either use actor(s) or subsystems

[crystal hatch]

there's no (easy) way to have a single AI controller control multiple pawns. The design is One Pawn <-> One Controller.

[sudden citrus]

@crystal hatch would Gameplay Abilities Behaviors reasonably serve as building blocks for an AI system where there's a separate decision making layer on top of BTs/FSMs/whatever?

[crystal hatch]

You can make Gameplay Abilities work pretty well with AI (see Paragon bots). One problem with GAs is that there's so many different ways to implement things, so much freedom, that if you don't define clear rules for the designers building abilities you'll have hard time having AI reason about the available abilities and their effects.

separate decision making layer on top of BTs/FSMs/whatever?
can you elaborate on this part?

[sudden citrus]

For example a utility-based system which has a few BTs to choose from, and runs one in a loop or once, etc. It seems to be a popular pattern

OH sorry I said Gameplay Abilities instead of Gameplay Behaviors. I meant Gameplay Behaviors

[real dock]

I have a simple actor MeshActor with only a skeletal mesh component. If I tell my AI to Controller->MoveToActor(MeshActor) it works great.

I have an exact duplicate of MeshActor with the only difference being its root component is a Scene component rather than the static mesh actor. Let's call this SceneActor.

If I tell the same exact AI to Controller->MoveToActor(SceneActor) it does move to the correct location, but when it arrives, it stands there for 5 seconds and then the path following request result ends with a Blocked error.

In some cases it reports failure (blocked) even though it is WITHIN the acceptance radius of the original request!

Something about the scene component in the destination actor is causing the navigation system to completely break.

Any ideas what I should look at?

[uneven cloud]

Unless you are writing a lot of expensive functionality, the AI Controller is not going to be what hurts perf wise with ~50 AI. It's going to be the character movement component.

[uneven cloud]

The better option would be to use run behavior dynamic where the tree is set dynamicly based on utility.

[crystal hatch]

The location PathFollowingComponent is testing against is different for these two setups. I suggest using a larger acceptance radius and if that doesn't do it debugging UPathFollowingComponent::HasReachedInternal - it's usually the height test that fails in such cases.

[sudden citrus]

Not really because then you're tied on the blackboard asset. Five years later you need to search for a BB key with a machete

[real dock]

Thank you for the insight. I will give it a shot.

[charred lava]

see Paragon bots where? 👀

[crystal hatch]

youtube now, I guess 🙂 Also:
http://www.gameaipro.com/GameAIPro3/GameAIPro3_Chapter24_Being_Where_It_Counts_Telling_Paragon_Bots_Where_to_Go.pdf
http://www.gameaipro.com/GameAIPro3/GameAIPro3_Chapter32_Paragon_Bots_A_Bag_of_Tricks.pdf

[charred lava]

okay, thanks 😄

[uneven cloud]

Only if you make super specific keys. The AAA game I'm working on has at most 10 BB keys.

[sudden citrus]

I'm architecting against other people's mistakes 🙂

[uneven cloud]

You think I don't work with designers? It's better to teach than to architect for their mistakes.

[sudden citrus]

It's a neat system all in all - I've got behavior trees wrapped in objects which handle the initialisation of the BB and can be parametrised by other systems thanks to that. There is some overhead but it feels clean.

[sudden citrus]

That is true, the way I see it I'm saving myself some technical teaching so that we can focus on the interesting stuff

[uneven cloud]

And I see it as getting better AI design by people who understand how AI functions.

[sudden citrus]

Do you have dedicated AI designers?

And @crystal hatch so would Gameplay Behaviours be a good fit for such a system? I'm guessing yes but wanted to doublecheck the intent since they first appear for Smart Object use

[crystal hatch]

In principle it could work, but I have no experience in that regard.

[uneven cloud]

I have 1 tech designer dedicated to combat AI, but other designers make AI.

I've done it the way you are suggesting on Psychonauts 2. I didn't use the gameplay behavior, but used the same functionality that it does for starting/ending a tree. It made things incredibly difficult to create new AI.

[sudden citrus]

What were the issues? If you are allowed to say

[uneven cloud]

The issues were that we had to initialize every blackboard key and clear it when the utility state changed. It was a lot of extra work and it created a lot of bugs.

You'd think it would be worse with sharing a BB, but I find people to be more intentional with their BB keys.

[sudden citrus]

Do I have to worry about it if I bin the whole BB every time I pick a new behaviour?

[celest python]

When (or "why" i guess) designers work on AI related systems? So far in the teams I worked, I always tried to make them not touch anything I do. Because one way or another even most simplest behavior was too complex for a designer to make changes, unless its just editing some properties in the details tab.

I wonder what designers do in bigger teams

for example do they ever script with BP? - in AI context -

[sudden citrus]

Just one example but Rocksteady is very code driven on AI side, almost to the level of hardcoding behaviour

[crystal hatch]

even most simplest behavior was too complex for a designer
Daym, you're working with the wrong people! 🙂
One of the things I love seeing is what those pesky-yet-creative non-programmers do with the tool we provide them with. Every team I worked with in the past had people like that that just needed right tools and some trust from the control-hogging programmers 😉

[celest python]

Daym, you're working with the wrong people!
🥲

[crystal hatch]

Yes. One of my favorite tech designers at Epic does all sorts of crazy behavior prototypes spitting out BT nodes and whole trees like it was nothing. Some of it sticks, some not, and the ones that do and are not performant get converted to native. Some are even good enough in BP!

[celest python]

I'm jealous now 😄 I'm too used to also designing what I created 😅

[sudden citrus]

More and more non-technical people join games, and that's good. I'll soon have 3 paper-only designers on the team. The further this shifts, the more an engineer becomes a system builder rather than content builder

[uneven cloud]

As an AI programmer, you have to be a designer as well. That also means making good tools that reduces the complication of making good AI. I've worked on teams where no designer would touch AI and ones that they felt empowered to. It's always better when more people can try interesting things.

[elfin hinge]

I though so too, but I have expensive functionality it seems. Because when I increase from 20 to 50, my frames drop from 60 to 40~. Models are Paragon regular minions, Behaviour Tree got 3 EQS (1 EQS at root and 2 at different branches) and I've got GAS. With SizeMap I can see that one AI has 200mbs and I've noticed with stats unit I have Game and GPU 26ms usage. So I'm just trying to figure out is it really that expensive or is it because I have seperate AI Controllers for each of those enemies (50 enemies, 50 controllers of the same Class). But thanks for response.

[uneven cloud]

That sounds like you need to profile it.

[crystal hatch]

I have Game and GPU 26ms usage.
this can indicate it's the meshes/animation cost, not AI per se.

[crystal hatch]

Also, remember Paragon? Every minion had its own AIController. The number of AI controllers is not an issue, what they do might be.

[elfin hinge]

Yeah when I turn off behaviour tree and leave AI controllers, it's not the behaviour actions for sure, seems like Paragon mobs is the issue... Thanks both of you.

[chilly bolt]

Navmesh doesnt work in Shipping build
Hello there, has anyone ever faced the problem with navmesh rebuild in shipping build? In development/Debug everything works fine, but when we tried shipping build, we would get super weird behaviour like whole navmesh is moved somewhere, it doesnt rebuild, etc etc.. I read some discussions with same description but no luck with a solution so far

[crystal hatch]

Navmesh doesnt work in Shipping build
Works on my machine 😄
super weird behaviour like whole navmesh is moved somewhere
ensure your navmesh actor instance location is (0,0,0), or even better: remove it and let the navigation system recreate it. And don't forget to save your map 😄

[chilly bolt]

what a joker nah 🙂
c'mon man, there bunch of links on the internet about the same issue. e.g
https://forums.unrealengine.com/t/ai-navigation-works-on-development-build-but-not-shipping/396802
or
https://forums.unrealengine.com/t/navmesh-not-updating-on-shipping-build/466418

so have you ever heard about something like that.. we are in a hurry a bit.. We need to deliver the build to a Distributor tonight but till today no one bother to try shipping build and now we have a real deal in here 😦

[crystal hatch]

have you tried turning it off an on again (delete navmesh and auto recreate)?

[uneven cloud]

You should always be testing a shipping or test build.

Deleting the nav mesh actor and force rebuild typically fixes things.

If it doesn't, you can try turning on dynamic nav regen and turn on force build on load.

[lyric flint]

I am having some issues getting my AI characters to hear a grenade sound. It worked a few days ago, but now it doesn't, and I can't for the life of me figure it out. Can someone please help me?

[chilly bolt]

that's not the case.. we generate navmesh on the fly. On a mission start, when an invisible obstacle(like quest barriers) disappears, etc etc.. everything runs smooth on development build but since tried the shipping build, things tend to be quite complicated. Navmesh is still there, but slightly moved. It is really really weird

[crystal hatch]

that's not the case
what's not the case? Even if your navmesh is fully dynamic there's a RecastNavMesh instance in your map (or your local equivalent; that's assuming you're not doing anything crazy with your navmesh setup). I suggest opening up your persistent map, deleting navigation data actors and rebuilding. Give it a try and report back.

[ocean wren]

Had a great meeting with an immersive arts collective I'm working with on a story generation toolset last week. Basically brainstorming the toolset they wanted for immersive storytelling. I kind of miss that design discussion, just doesn't tend to happen in academia. They had some whacky ideas and we're going to try some of them. Will have a prototype plugin working before xmas and some fun demo videos early next year (its part of the funding agreement that the company demo's their use of the toolset). Looking forward to sharing some of it (things like voice controlled AI scene generation stuff)

My design notepad from the meeting looks kind of weird too.. got words like "chaos field" and the like 🙂

voice controlled particle effects... hahahaha.. just thinking about it is funny 😉

natural language is quite hot as a topic for generative stuff right now. Recommend looking at it. Huggingface do some really nice usable libraries for it too.

[misty gale]

Voice controlled sounds pretty awesome

I had an idea using that for VR DnD sessions

Procedurally generated as the narrator tells the story and sets the scene

[real dock]

Just wanted to circle back and thank you for your help. You were exactly right about the problem.

I added an override of HasReachedDestination with some visual debugging aids and indeed the goal was hovering above the ground. Unrelated bug caused that, everything now fixed, thanks a ton for sharing your insight. 👍

[severe cypress]

My AI is a massive amount of the data going to clients on my dedicated server, specifically the ReplicatedMovement. Is there a good way to clean this up / optimize without making the AI movement choppy?

[crystal hatch]

There's no substitute for experience - I've seen this issue a dozen of times 🙂

[safe crescent]

Anybody know why my navmesh seems to be in the sky?

No aimoveto or moving nodes seem to work

[crystal hatch]

yeah. The navmesh instance location is borked. Quickest fix is to delete navmesh actor and rebuild navigation.

[safe crescent]

Alright i'll try that

Seems to appear fine when not in play mode though

which is the weird part

[safe crescent]

Still got the same problem

[crystal hatch]

I noticed you're using World Partitions. TBH I don't have much experience with navmesh in this context. The only way I can suggest is debugging the relevant code, sorry.

[safe crescent]

All good thanks for you help

could it perhaps be something to do with the size of the navmesh since i have a pretty large map right now

[crystal hatch]

I do recall there's a commandlet for some navigation world building for partitioned worlds. Have you tried that?

[safe crescent]

Nope i havent but i'll try find it

[crystal hatch]

I don't think so. How large is "pretty large"?

[safe crescent]

The size of the navmesh is like 3250x3250x50

[crystal hatch]

3250 what? Tiles? Meters?

[safe crescent]

I think meteres

[crystal hatch]

it's not large 😉

[safe crescent]

Just like the scale of the navmesh is 3250x3250 under the component like transform rot and scale section

[safe crescent]

Alright

Could be something to do with the engine version since im on 5.1 preview 2

[crystal hatch]

umm... what? You have non-identity scale for your navmesh? The RecastNavMesh actor (or the equivalent)?

[safe crescent]

The navmesh boybds volume

bounds*

[crystal hatch]

right. So it's 3250 times the default volume size 🙂 No idea what that is 🙂 Around a meter I guess. Still - not large.

ok, so when I said remove the navmesh actor did you remove the bounds or RecastNavMesh actor?

[safe crescent]

I tried both

[safe crescent]

I tried changing the position of the navmesh boudns up by 1 then rebuilding navigation and it just kinda works now

Quite weird but hey atleast its fixed

[serene basin]

The AI perception callbacks only seem to work when i walk into sight radius, if the actor spawns next to me that event doesnt get triggered? Anyone knows how to solve this with the AI perception component ?

[serene basin]

Ok I seemed to solve it by just disabling and enabling the component

[proven plinth]

I've created a BTService that on an interval attempts to attack. However when that service executes it hard locks the UE editor and i have to force quit and re-open UE. I've reduced everything inside of it down to a is valid check and a print string trying to trouble shoot it. I was modeling my BT after Lyra's BT_Lyra_Shooter_Bot example, that made use of a BTS_Shoot. Is there a catch/known issue around services with intervals? I get 0 logs, and cannot tell what is going on.

This is the BT if it helps at all.

[crystal hatch]

I'd prefer to see the service implementation.

[proven plinth]

When debugging, it makes it to print string, and hitting next is when it locks up. (its not specifically the print string, just whatever was the last node in the service)

[celest python]

so this 100% happens right after that service executed?

[proven plinth]

yes, as soon as the interval is triggered. when i bump it up to say, 5s it will move towards its target for 5s then lock up. I'm hitting play PIE if it matters?

removing the service, no more locking, NPC will just move to its target and then stick to it like glue..

[celest python]

is there any chance you are hitting an ensure() or something similar in C++ side

because assertions collects callstacks when you hit them if you are not in debug mode

and it takes... quite a lot of time

did you try waiting 2-3 minutes?

you would see it in logs i guess though but.. nothing else comes to my mind

[proven plinth]

So, once i started seeing this, I switched to actively debugging in Rider, i never get a breakpoint from an exception, and ya.. nothing in logs. I'll let it sit here for ~5 and see if i get anything different.

[celest python]

if you ran with DebugGame Editor on Rider it should open the IDE immediately

if it didnt, then something else is wrong

[proven plinth]

I'm running Debug with the Development Editor | Win64 configuration

i'll switch to DebugGame and see what i get

[celest python]

it wont change anything if you're 100% sure debugger is attached to your editor, which should be if you ran with Debug

[proven plinth]

it is. i hit other breakpoints

[celest python]

can you try printing callstack on rider debugger

one sec

#cpp message

do this when editor is frozen

see what is evaluated/executed last time

[proven plinth]

so, hit pause and then invoke that? Because you dont get an immediate window unless you are at a breakpoint (that i'm aware of)

[celest python]

immediate window can be opened manually, at least in VS -- not sure if it allows you to do something with it tho

just play your game and let editor froze itself when you run the service

then invoke that

[proven plinth]

ahh that may have helped. I think I have a recursive call in c++ thats in an eternal loop.

Specifically the retrieval of the ASC in that service.

[celest python]

interesting

but nice, if it helped

[spring jetty]

is it possible to start running a BT from a specific branch ? or maybe go back to a previous task ? I have reached a point that I need to do sth like this. is it possible without cluttering the BehaviorTree with more decorators or services.. etc

[uneven cloud]

I really wouldn't recommend using the BT for conversation and certainly not to drive the player. There's the experimental common conversation plugin that you can try.

[spring jetty]

I have thought about doing it in code and not BT but it might get ugly, I thought this might be more intuitive but didn't think, implementing a back functionality would be hard

[celest python]

I cant recommend this enough: https://github.com/MothCocoon/FlowGraph/

[spring jetty]

I dont' like using plugins, I'd rather implement sth myself and learn sth new, how would you implement a dialogSystem then >?

[celest python]

You can learn from plugins too, though 😉

FlowGraph is basically a generic graph that lets you run latent nodes and it lets you customize them very easily

so its design makes creating dialogue system easier by its nature

for other type of systems there are GDCs for in-depth explanations

mostly for branching dialogue systems like in RPG games

also this one: https://youtu.be/tAbBID3N64A -- which is my favorite because very flexible

[spring jetty]

I see thx, what is GDC though ?

[celest python]

Game Developers Conference

developers from industry do talks and GDC organization releases videos/slideshows

https://www.youtube.com/c/Gdconf

[spring jetty]

oh, thx for sharing!

[uneven cloud]

Try out common conversation. It has an editor tool that is similar to the BT, but works for conversation.

[drowsy plinth]

how to prevent AI from getting dragged on edges?

he is getting temporarily stuck on corners

[uneven cloud]

Make sure the capsule fits the character correctly and the agent settings (in project settings) fit the capsule.

[drowsy plinth]

i changed to different ai parent, crowd something

and ai stopped catching corners.

[normal mist]

What is the difference between a USmartObjectBehaviorDefinition and a UGameplayBehavior? It seems the UGameplayBehavior also has access to the smart object owner.

wait a minute. the rabbit hole just got deeper.

what is the difference between gameplay behaviors and gameplay interactions?

[severe cypress]

What kind of movement component do games like Ark use for AI?

[normal mist]

I notice that teams tend to be stored in the PlayerState (looking at ShooterGame). If I have a structure/building that can be assigned a team (say, a barracks for auto spawning pawns of its team), does that mean I should have the barracks be a pawn with an AIController that has a PlayerState?

feels a little overkill to make a structure be possessed by an AIController, however I don't want to have two different implementations of interacting with an Actor's teams

[misty wharf]

This depends on how you intend it to be identified

Nothing says you can't simply have a field on your building that tells you which team it belongs to

PlayerState etc. would be used for things that are players (human or AI), of course you don't need to do all that for things that aren't

[normal mist]

Thanks, I appreciate the response. I'm using gameplay behaviors & smart objects to facilitate a "capture" interaction. Feasibly this could be used on buildings or units. If the implementation of the actor's team is the same, then the gameplay behavior is streamlined. If for units its stored in their playerstate, and for structures its in their actor, then the implementation for team-checking becomes divergent.

Also, does anyone know the difference between User Tags & Object Tags for a smart object definition?

I'm reading through the source code and looking at its usage in the USmartObjectSubsystem but I'm still scratching my head

[misty wharf]

I'm not familiar with the smart object stuff, but you should be able to identify teams by using IGenericTeamAgentInterface

So regardless of how the information is stored under the hood, if the actor implements that, that should be enough.... or if not, guess you could always use a custom interface :)

[cosmic gulch]

so say part of my behavior tree is devoted to roaming. Is adding a "Is roaming enabled decorator" then changing the value to false on child classes the most elegant way to handle that for child classes?

[tidal spade]

thanks for posting this! I just watched through it and it's fascinating. I'd absolutely love to play around with this sort of system, but it's well beyond my capabilities to implement something like this on my own. do you know of any marketplace plugins that implement something similar?

[uneven cloud]

The USmartObejectBehaviorDefinition is just a wrapper for a gameplay behavior, used by smart objects to drive behavior.

[crystal hatch]

(looking at ShooterGame)
Don't. Shooter game is extremely outdated. Use Lyra instead.

[normal mist]

Thanks for the response @crystal hatch

[lethal helm]

any idea why a dynamically spawned smart link isn't functioning

dynamic regular links work but not smart

okay... they spontenously started working after a editor restart

[celest python]

there isnt one, but I'm developing one to release as open source.. though I'm notorious with releasing things once and never updating it 😄 I'm also bloated with tasks in my daily job so not sure when I can finish it.

actually I finished it a while ago already but I wanted to provide a custom parser to let people write conditions like how it's presented in that GDC.. like someValue > 4 && otherValue == true on editor.. probably with csv or json support. but it was more time consuming than I thought and I was out of free time to continue developing the plugin

a simpler approach is tag based one, but only useful for "barks" rather than "conversations". an example from Division 2 is on page 67

and there is an example in CryEngine, that is more advanced than anything I've seen as an example of what explained in GDC but since it was made in CryEngine framework I couldnt reason many things written: https://github.com/ValtoGameEngines/CryEngine/tree/release/Code/CryEngine/CryDynamicResponseSystem

[forest tundra]

I said if the NPC doesn't see us, it should wander around randomly, if it sees us, it should run towards us. I adjusted the NPC tree by checking it with the bool variable. but why does the NPC stop when it sees us? can someone help?

[crystal hatch]

if you're referring to something said or posted in the past please reply to it, otherwise we're missing context

[tidal spade]

that's a super cool project you've been working on then! do you have a repository or anything like that I could follow in case you ever decide to post your code?

the tag-based approach is an interesting setup. I'd definitely want to lean heavily on gameplay tags but would also want to utilize other things like attributes from the gameplay ability system. I've been looking for the video of the presentation from that slide show but it seems the archives are incomplete

[celest python]

GDC released only PDF for that, video might be released in the future or maybe its paywalled

[celest python]

Sadly there isnt one yet, even though core functionality exists I still need to polish it and go through some tests. But I'll definitely share it on github when it's done

The "fuzzy tuple/pattern matching" they mention in the GDC is actually a "language feature", and C++ doesnt implement it by default - they wrote a custom software/engine for it to make it optimized afaik. Its probably implemented in some other old school languages like haskell, erlang.. and maaaybe in C#. I went through tons of research about finding the best and most optimized way to implement it on my own, but I ended up in an alternative solution. But I couldnt test it yet, so even if I release it today it'll be unusable because there wont be a proper way to add dialogues because you'll be adding structs to nested arrays in editor, which is infeasible to work with if you are developing something serious 😄

thats why I'm working on that parser I mentioned earlier, so designers will be able to type to a string field in editor and system will convert it into actual data with reflection system

[tidal spade]

oh wow. I knew the solution was over my head, but that's a lot more involved than what I expected.

I just followed you on github so hopefully I'll see if you ever decide to publish any of your work. thanks a lot 🙂

[finite parrot]

hello, I'm new to ai behavior tree

what does lock ai logic mean? does it mean during this task, BT stops checking conditions?

[celest python]

animation system and AI system (brain component) has resource locking systems

it pauses the execution logic when you lock

so lets say if you are already running MoveTo task in BT, and you run that BP node, it stops the execution of BT so two moveto nodes wont conflict

or any other behavior

[finite parrot]

What if I still want to check BT condition during this move? do I need to store this task and manually kill it?

[celest python]

just uncheck the tickbox

if you dont lock it, it will continue evaluating the tree

[finite parrot]

yes but will this task still be running?

if condition does not satisfy

[celest python]

which task and which condition?

are you running this inside of a BTTask

[finite parrot]

yes it's inside a bttask

[celest python]

path following logic is not a part of behavior tree, so even without BT it would work if you just manage to call Move To (its actually a function in AIController, MoveTo and the node you posted are just wrappers for it with additional params to make your life easier)

if you are in BTTask, do not use Lock AI Logic

if you abort the current task or switch to another task, call StopMovement in AIController

[finite parrot]

ok

thanks

Solves my problem. Want my AI to break circling player task and switch to chasing task immediately if player gets out of attack range when the circling move is still running.

Adding a little bit of cooldown to the branch seems to help with BT flashing too fast.

[leaden fulcrum]

I struggle to understand the condition for Bool Match on a Trace test in environment query:
If I want to get the point of a donut from which the player is visible then it seems like I need to set Bool Match to false. Which doesn't make sense to me, since in the block it displays it as "require not hit".
How do I need to understand bool match?

From the docs:

When performing a Test, not matching this value will not change the score.
I interpret this as all points that don't lead to a sucessfuly trace (line of sight) should be filtered out. Therefore, Bool Match should be true, since I only want to keep the points where the trace was indeed sucessful. However, something seems to be wrong about my logic here, because to get it to work I need to set Bool Match to false.

[spring inlet]

is there some way to integrate something like exchangeable sub routines in the AI, or to execute two BTs for one actor?

[keen crow]

Maybe you are looking for Run Dy namic behavior task in BT?

[celest python]

What is the best way of checking line of sight without using perception component?

My AIs are lightweight and doesnt require anything perception component provides other than checking LoS

I guess there is AIController::LineOfSightTo

but it also seems to be doing linetraces which makes me think it its really accurate or not

[lusty nebula]

what else would you expect the system to use if not a line trace?

[celest python]

It does 4 line traces to different locations, I'm wondering how it detects things between those traces

I think I'll write a custom one..

[uneven cloud]

What makes you think it's not really accurate? It's a lot better once you adjust the view location, but what kind of precision do you actually need?

[celest python]

Since there are only four (+ trace from view point makes it five) traces and I thought if radius is big enough traces might not be able to hit actors between

[uneven cloud]

Have you tried it? The vast majority of the time you don't actually need that complicated of line of sight functionality.

[ocean wren]

Oh wow, that was embarassing, thought I'd goofed with some rendering because it was running slowly.. turns out I was literally downloading the planet in the background, forgot about those torrent files 🙂 hahaha

[ocean wren]

The stalker series picked random bones and just did traces to those.. I did the same and it felt pretty decent.. perhaps not 100% immediate every time, but pretty damn near.. and you can priorities specific bones if you like

[spring inlet]

thank you, that seems to work in combination with simple parallel 🙂

[stoic cliff]

Hi, so im quite new to AI but saw it as an inevitable step in the game im making. So, im making an rts style game, where units can be selected and pressed to move. Currently, i can make them move with the Move To Location command, but they all try to move to same position. How can i utilize the AI systems to make them move to form formations automatically and space out? And select optimal location based on current position?

And how can i program them to move out of the way when another unit is attempting to move through them?

Or should that just be accomplished by allowing the moving actor to push the others? 🤔

[celest python]

yeah, it keeps returning true

[celest python]

easy mode: using EQS and trying to move on with what it can provide
difficult mode: writing a path following component child and making AIs follow given path with formations, with avoidance and such

[stoic cliff]

I just saw that EQS was a feature in development, and i want to avoid bugs as much as possible. So i think i will try the path following method. Thanks!

[crystal hatch]

EQS is marked as experimental only due to the tool’s UI, which is not to Epic’s standards and we don’t have time to work on it. The core of EQS is as solid as it gets and there’s a countless number of games that shipped with it (both internal end external, including all Epic’s games from past 8ish years). The core is so mature no more than two issues get reported a year for it, and it’s edge cases usually.

[stoic cliff]

Ah okay, thank you. I'm still just learning about AI, and was thinking to stick to BT's and BB's to make sure i don't confuse myself with some special feature that Epic is working on :)
But if it is so solid, i will use it.

[crystal hatch]

EQS also ties well with BT and BB, and it's super powerful for a lot of things (not only in-game AI). I strongly encourage everyone to get to know it better and use it to its advantage.
Re Epic still working on it - I'm actually spending some (rare) spare cycles on planning EQS2.0 actually, based on Mass 😄

this might not have been announced publicly yet, so.. there you go 😄

[stoic cliff]

Sounds exciting! :)
Thank you

[haughty coral]

I wonder how EQS would compare to this https://twitter.com/skylonxe/status/1588861427380805633?s=46&t=UkmCzsMlxlj7ukCZRwzgIg . To me it looks a bit too much like EQS. Not sure what was the motivation to create this instead of using EQS. Looking forward to any movement in EQS, I hope it happens one day. My wishes are only more modern handling of stuff like param types (I guess planned as I heard state tree has that already covered). Now I also have a great opportunity to say something I imagined to say at the beginning of development of all AI systems: don’t you dare to use there somewhere AAIController instead of AActor 😆 These things are too useful to be tied to AI stuff.

Ondrej Hrušovský (@Skylonxe)

Generic targeting system plugin was recently added to UE5. It can be used also with abilities to build your targeting in a modular way using tasks like selections, filters and sorts. https://t.co/VVHuApxU4B

Twitter · Nov 5, 2022, 11:50 AM

[crystal hatch]

My wishes are only more modern handling of stuff like param types (I guess planned as I heard state tree has that already covered)
yep.

don’t you dare to use there somewhere AAIController instead of AActor 😆
I do what I want! 😛 But yeah, that's the plan. Overreliance on the AI controller is a mistake of my long gone AI programming youth 😉

[haughty coral]

Awesome :P thanks for the reply.

[celest python]

actual RTS games require path following method.. but a different one of course, because the more your scale increases the more work and optimization you have to do. Unreal is not suitable for big scaled RTS games

For example unreal's current actor transform system was used by Sony for PS2, but they refactored it at 2006 for PS3, and at 2007 they made a talk about how terrible it was. And unreal still using that system in 2022. dont expect to have moving 1k actors in Unreal without some crazy changes in the core system of UE 😄

[crystal hatch]

Mass aims to address that without actually gutting the engine.

[celest python]

yeah, to be honest, it was the best idea so far I've seen 😄

it literally provides an alternate gameplay framework

[crystal hatch]

not yet, but we do want to provide a lot

[celest python]

now thinking about it, since years looks like AI team actually had the lead of development for most of the gameplay framework in Unreal Engine

so I'm not surprised #mass came out from AI team 😄

[crystal hatch]

That's not fair to the rest of the Framework team. They do a lot of grunt work, and do important lower-level changes to support open world/large scale games. It's just that those features are less visible on the surface.

[celest python]

Yeah, sure. Though when developing a project without diving into low-level code you mostly see things are shaped by AI team

[misty gale]

i cant see anything blocking the line of sight here tho, assuming the target is the cube

sorta sounds like it just traces towards target, regardless of source rotation

[celest python]

yeah i guess its not what I'm looking for

until I write my custom one I'll go with that function + dot check i guess

[ocean wren]

Avoidance Alice wants a word

[stoic cliff]

I think it will be fine 😄
Im not planning on doing all of it alone, so later down the road i can always get help to develop more advanced systems. For now i found a workaround which makes my units move a bit like they do in games like starcraft, which is decent.

[celest python]

Am I too young or green to understand this joke?

[celest python]

if we put performance and technical worries to a side, difference between EQS and pathfinding/avoidance choice is, EQS lets you choose a point in a single place, meanwhile other option lets you manipulate the currently followed path

so its more flexible for a fast paced action rts where a lot of things going on in the screen

[ocean wren]

So, had a few days of working in Unity.. I like the C# language, but got really annoyed at the pass by reference nature of its data structures.. real pain in the arse

Kept blowing past my geometry buffers because it was using a class for primitives and apparently it was treating every geometry primitive as a shared reference rather than what I actually wanted (individual values). Sheesh

Now back to voice control plugin

[celest python]

aww

[clever mason]

I'm on the verge of rage quitting! How do I create a [Combat Manager].

I'm trying to create an open world game similar to Shadow of War.

I need Enemy AI to attack one at a time and I keep failing. Pictures would help pls.

[celest python]

are you trying to have something like a token system?

only single AI attacks to player at once

[celest python]

https://youtu.be/2KQNpQD8Ayo?t=2483

[clever mason]

@celest python I watched the important bits but I still don't get it. Do you have blueprints or a behavior tree I could follow to at least get step 1

[celest python]

I have an open source plugin for that: https://github.com/intaxwashere/AITokenSystemUE

all you need to do is setup the plugin properly then subclass UAITokenBase and override SelectBestSubscriber function and do your own logic to select best one

[ocean wren]

OPEN SOURCE!

goddam it 🙂

taking our jobs etc..

[celest python]

taking our jobs
I just need.. uh.. checks notes have 20+ years of CS experience and some god-level ML knowledge to do that.. 😂

[ocean wren]

Yeah, well, easy to come by 🙂 hhahaha

[celest python]

See you 20 years later
unless I pass the exams to Lincoln University to study CS 😂

[ocean wren]

I think you'd be ok.. given you're already programming

You can learn a hell of a lot by actually doing it 🙂

[celest python]

One of the most solid computer scientist I've spoke earlier was said "the useful part was about studying CS is knowing what book to read and the homework assignments" and "rest of the time was wasted"

so if I'd ever study CS, it would be probably only to involve into ML researches or projects etc.

because in 2022 you can already access information everywhere and get code reviews, open source projects everywhere etc.

[ocean wren]

Yeah, its a bit of a weird viewpoint, but yes you can pretty much access anything

degrees are more like social clubs than anything really important knowledge wise

BUT, its a friendly place to learn stuff if you're of a mind to, which isn't that often 🙂

I mean, I've got a masters student looking at neural radiance fields to try and generate usable meshes for them.. I'd argue that was a good thing to work on and study

The reality is, most students would be happy just paying the $$ and taking the degree paper.. I'd rather we offered them for sale, but apparently we can't 🙂

The whole system is messed up

[celest python]

I'd rather we offered them for sale, but apparently we can't
😄

[ocean wren]

Would be better honestly, to make a small studio and just work on really hard stuff for a few years 🙂

[celest python]

I guess that's what you're turning your classroom into, right? because whenever I see you are talking about your project your students also always involved in it - unless its completely commercial thing

[ocean wren]

That said, there's a lot to learn out there.. if you're that way inclined.. lots of avenues for mindful approaches (neuroscience and neuroaesthetics for instance)

I wish I could do more

mostly I'm just doing commercial contracts myself.. don't have enough time to do enough of it to get a proper business going

So much opportunity, so little time 🙂

[clever mason]

@celest python Wait I never used github before does that mean I have to install Git to UE4 plug-in or can I just use the code immediately?

[celest python]

I guess you want to continue in academia too, then?

[celest python]

you can either use GitHub desktop or just download as zip directly

then paste the folder to your ProjectFolder/Plugins

then compile your project

[ocean wren]

Honestly, not even sure anymore.. we're about to go on strike, I suspect it'll get nasty, in some ways I'm hoping I get laid off or something, will give me the incentive to take a new approach

[celest python]

if you are not using C++ you can use this to make UE automatically compile the plugin: https://landelare.github.io/2022/09/27/tips-and-tricks.html#automatically-update-c-binaries

[ocean wren]

I can come back to the research later.. commercializing it would be fun right now

[celest python]

that wasnt the thing I expected to hear 😅

[ocean wren]

Lots of weird stuff going on in the UK right now

[celest python]

but yeah no doubt there is a valid reason behind it

[ocean wren]

and education is no exception

[celest python]

my country is no different so I completely understand

[ocean wren]

They're basically pushing us to bring in $$ as consultancy etc.. at which point I'd rather work for myself 🙂

[celest python]

that'd be way better for you I guess 😄

who knows maybe you can even create some cool stuff for UE too 😄

[ocean wren]

Less dealing with beaurocracy at least

Yeah, I've already got a big company in London interested in my current project's outcome

After Thursdays event 🙂

I guess voice based AI driven pcg got them a bit interested

[celest python]

are you thinking about some end-user facing or something like a (social) platform related AI work soon?

be it something related with game dev workflow or something like text-generator AI for users etc

[ocean wren]

This is mostly production tooling

Bit in the middle.. the use case is non-expert creative types in VR/AR

and virtual production

Basically, its all immersive Unreal Engine stuff, but we're exploring voice driven PCG as a toolset for creative exploration in that space..

given that we've got pretty solid natural language tools and voice understanding and voice generation.. and good language modelling of characters and the like..

I mean anyone can do this stuff, as long as they grok ML and PCG and UE 🙂 hahahaha..

[celest python]

I guess industry relies on a few open source frameworks.. 😅

[ocean wren]

I've got some second year students looking at maybe tying gpt3 and stable diffusion together

there's so much stuff to sort of smoosh together if you're of a mind to

[celest python]

I guess exact same thing introduced to beta.character.ai a few days ago

not SD but a different type of it

[ocean wren]

yeah, I told them to look at that too

those large language models are pretty much everywhere at this point

[celest python]

if everything shares the same base framework where is the innovation tho

everything is GPT3

[ocean wren]

its not all the same base.. there's bloom which is an open source one, there's gpt3 which is OpenAI + Microsoft, there's T5XXL which is googles one, there are a few more. There's huggingface's api etc.

and there are like a dozen more being trained at massive scale currently, both open and closed source

[celest python]

what it takes to create one from scratch? -- if you dont mind explaining like I dont know anything 😅

is it same thing like not creating a game engine from zero but working on UE

[ocean wren]

basically, scrape every bit of text you can from the internet 🙂

then train a model with attention across all of it for thousands of hours 🙂

[normal mist]

What is the difference between User Tag Filter and Object Activation Tag Filter in a SmartObject Definition?

I looked at the source code and couldn't quite infer

[celest python]

I guess the more computation power the more faster it is -- so its mostly $$ stopping people to train GPT3s from zero

[ocean wren]

the basic idea of these language models, is that at larger scale, they learn concepts.. not just words like "dog" or "cat", but they have the concept of dog as a cluster in the language model.. so dog, chihuahua, canine etc.. all live in the same area in latent space (like a reduced set of numbers that captures the core of language)

they're spending like half a million quid training these models

or more

[celest python]

quid?

[ocean wren]

GBP 🙂

[celest python]

oh 😄

[ocean wren]

pounds

the power of the models, is that you can ask them to do things like fill-in-the-blanks on masked input.. "the ??? sat on the mat" style

or you can ask them to give you all the verbs, or nouns

you can ask them questions about the sentences

you can see longer term relationships between texts, because they've literally seen all the text there is on a concept

[celest python]

how it "interprets" the text? I noticed beta.character.ai first tries to interpret the text you send then stores "facts" to its memory to build up the next text its going to generate. Ofc it must be something very complex on low level side but on high level idea it doesnt seem like rocket science

[ocean wren]

so for instance your character.ai thing, they've seen all the albert einstien text out there.. his history, his writing, people discussing what he said etc..

it interprets it as a latent vector... its basically a compressed code (bunch of floats)

the point is that you can give it something like an image that has been compressed into its own latent vector space... and something like Clip, which was trained on image/text pairs.. so the text gets compressed into a vector along with the image.. they map to the same space in vector space

[celest python]

so there is a "nativization" step to make AI approximate the output

[ocean wren]

so for instance, the word "dog" is in the same latent space as the image of a dog

[celest python]

turns things into "nueral" language

[ocean wren]

its a tensor of numbers which represents a notional concept space, where different words/images/whatever are expressed as clusters of similar things

i.e. dog image maps to same space as "dog" text

(not quite, but conceptually)

[celest python]

sounds terrible for CPU cache, I guess I understand why GPUs are widely used 😄

[ocean wren]

from a practical point of view, they're just tokenized vectors

the point is, given billions of them.. the model learns the relationships between concepts

[celest python]

so if you just have.. enough power to supply the training you can just re-create a GPT3 with enough of engineering care?

[ocean wren]

i.e. it probably sees "hair" quite often near the word "dog" so it "knows" that dogs are likely hairy

yeah, bloom is an open source gpt3, trained on a cluster..

[clever mason]

@celest python I feel like I'm walking in the dark. I've never used visual studio.

[celest python]

did you check the link I've posted?

if you put that to .ini file mentioned, it will compile automatically for you

so you dont have to deal with Visual Studio

or simply just open the .uproject and when it asks you to automatically compile things press yes

[celest python]

then mostly data scientist matter in here

to organize the data to not make AI biased to produce wrong results

so it can train itself

with less engineering care

so you dont have to manually train AI each time

[ocean wren]

There's a whole funnel of stuff to work on.. yeah, trying to unbias the training data is part of it

most of the time, you don't train AI from scratch.. you use a process called fine tuning, which basically freezes parts of the model and lets you train the other bits

with ML, you're basically pushing floats up and down in the model by tiny amounts

[celest python]

ughh ai looks so fun to work with

wish I could have time to work on ML

[ocean wren]

try some practical stuff and you'll be hooked

but be warned, you'll have to learn python (ptooey)

[celest python]

yeah switching to weak typed language from C++ wont be smooth i guess 😄 but a small price to pay

[celest python]

where should I start?

books i guess?

[ocean wren]

huggingface 🙂

[celest python]

I should directly start working with a framework?

[ocean wren]

hmm, books? nah not really.. try the fast.ai course, its free and pretty good to get straight into it

[celest python]

alright

thanks

[ocean wren]

you're very welcome

apparently they're doing an image generation course with fast.ai and jonathan whittaker.. due out in a week or so

the reason I mention fast.ai is that they basically go straight into doing something cool with ML, then back up and introduce all the parts.. its a neat way to teach

means you understand what it can do even though you don't know why until later 🙂

[clever mason]

@celest python Sorry I utterly failed to follow directions maybe I'll try something else.

[celest python]

if you can explain the steps you followed and why you failed I can try to help

[clever mason]

• Extracted the Token
• added the auto compile set up to UE4
• I didn't have projectfolder/plug-ins for my UE4 project so I created one in the actual project
• Nothing happened

[celest python]

can you show whats inside of your Plugins folder?

it should be Plugins/AnotherFolderThatYouCanNameToAnything/ExtractYourZipHere

UE looks down to paths by folders and finds .uplugins files on each folder to recognize them as plugins

if you directly extracted it to /Plugins it mgiht not work

[clever mason]

Alright I give it a shot after I take a break

[spring inlet]

is there a way to set the socket for the AI Perception component?

[cinder garnet]

having a problem with my AI, in the middle of a chase, or even if i stand in front of him, randomly he will stop chasing me and then start patrolling again. does anyone know what i should edit in what i have, or a way to fix this problem

right before i added his random patrolling, there was no problem with him following me

his peripheral vision angle shouldnt be the problem

[uneven cloud]

The way that you have it set up it will patrol regardless of if it's already chasing. Because you are doing it in tick (not recommended) and not checking if it's chasing.

[uneven cloud]

It's not a scene component, so you can't set a socket. You can override the functionality for getting the view location/rotation and use a specific socket.

[cinder garnet]

so what should i replace the event tick for

replaced with custom event, works now

[celest python]

you can recall event after this node

to make it loop

latent nodes and tick doesnt go well

[cinder garnet]

so what is it called, and do i link it to the is moving set?

and what is the noe called i have to replace with tick, still very next with this, only been doing it for a week

[celest python]

you can call the event node you created

the one you replaced with tick

to the is moving set?
after it

[cinder garnet]

Okay so delete the event tick, and put "IsMoving?" Where the event tick currently is

[meager imp]

I am having problems understanding what the different versions of "Abort" do
Abort Self / Abort Lower / Abort Both
any idea where I can get some info about this? 🤔

or help me understand this:

• Abort self is to cancel the subtree if the condition becomes false
• Abort lower is to cancel the other subtrees if the condition becomes true
  is this correct? 🤔

[crystal hatch]

Pretty much, yeah. "Abort lower" is short for "Abort Lower Priority branches", meaning if observed condition's value changes from false to true the gated branch will be reconsidered. If there are no valid tasks to run in this newly tested branch the currently running one will continue.

[severe tangle]

How can I set up Detour Crowd avoidance? I have my custom AI Controller set as AI Controller Class on a Character Blueprint, this custom AI Controller has Detour Crowd AIController as parent class, but it doesn't work. AI doesn't avoid other character..

[crystal hatch]

the other character needs to be part of detour simulation as well. If it's not an AI agent it needs a bit more work to get included in the sim. See class comment in CrowdManager.h

[severe tangle]

Oo I see, thank you!

[magic jasper]

Can you take the replication one too pls 😄 I think Mikko said it was on his plate when he took the BP one
https://github.com/EpicGames/UnrealEngine/pull/9280

[old lichen]

Isn't there a BT node for modifying BB values?
Do I need to create a task?

[celest python]

there isnt a generic one

[old lichen]

alright

BTW, I don't get the point of conditional loop decorator checking if a key is set or not rather than its value.
How can I do the same but checking a boolean value?

I mean, same as before? Do I create my own decorator?

Also I need the task to be executed at least once, like a do while

Can't see how to do that

[lyric flint]

Can AIPerception be used with an AI that can see 180 in horizontal and also vertical?
By default it seems that it can only see horizontal and a bit vertical as it seems a cylinder rather than spherical.

[plush cargo]

I have been trying to troubleshoot my knockback function and it seems that i cannot affect AI actors movement when they are running behaviour trees. Is there any easy way to override this?

[normal mist]

For some reason USmartObjectSubsystem::FindSmartObjects isn't finding any smart objects on a world partition level. However, it is finding smart objects on a non-world partition level

does anyone have any ideas?

I found the SmartObjectionColelction actor in the world partition level and hit "rebuild collection" and can see the smart objects in the collection

[uneven cloud]

How are you trying to knock back? Using launch character should still work.

[plush cargo]

checking now

[plush cargo]

this seems to work but isnt that because of how physics can affect the actor

[uneven cloud]

Launch character uses the movement component.

[plush cargo]

oh. well... launch character had a noticable effect but still something in the AI controller that interfered. Im trying first to interpolate the position based on the knockvector first.

but thanks for showing me launch character

[waxen hare]

Hello! I am trying to create some wall-traversing NPC's (Spider robots), but I'm only now realizing that navmeshes are Z-up only. Are there any plugins that does a decent job of vertical traversing, or will I just have to build my own pathfinding from scratch?

Haven't found anything premade so far

[uneven cloud]

Unless you added something, the only time the AI controller should be dealing with movement is when navigating.

[plush cargo]

there is a behaviour tree that controls the movement

if i just make a new actor and apply the knockback to it, it works. assigning the ai controller disables the knockback

i tried disable movement and stopping movement as tasks

nothing works

[uneven cloud]

The behavior tree only controls movement by navigating, unless you added something.

[plush cargo]

i will redo the BT and see where it stops working

[slow sphinx]

Man I have no idea what I did but suddenly my character won't move...I checked the player controller and the Simple Move To is firing...the character simply isn't moving....

[plush cargo]

This was the culprit. The ai guide i was following edited this value

[crystal hatch]

It's on Mikko, and it's going through internal discussions with network people (we the AI folk don't know everything. I know, it's hard to imagine 😉 ). The intent is definitely to pull it in, just needs massaging.

[crystal hatch]

SmartObjects do not properly support WP worlds just yet. I'm actually working on it right now.

[crystal hatch]

Check vislog (Default Answer #1!) and make sure navmesh is there.

[normal mist]

Thanks for the response. Would this explain why one of my smart objects returned in a FSmartObjectRequestResult not finding its proper SmartObjectComponent from USmartObjectSubsystem::GetSmartObjectComponentByRequestResult? I am not sure why thats happening.

Also, do you have a rough ETA on when they would support WP? Is this something I could submit a PR for?

[crystal hatch]

Would this explain why one of my smart objects returned in a FSmartObjectRequestResult not finding its proper SmartObjectComponent
This is not WP related, it's a feature of the Collection. The whole point of the collection is that you can simulate smart objects (using Mass under the hood) without needing the actor to be streamed in. This logic is a bit convoluted and I'm working on changing that.

Also, do you have a rough ETA on when they would support WP? Is this something I could submit a PR for?
This should be available in 5.2. If you want to use it asap (i.e. using Master branch) then it should be ready(ish) in around two weeks (baring no distractions). You can post PRs but note that it's not really an implementation issue, the architecture of the approach needs to change.

[normal mist]

Thanks so much for the info. Is there a way that I can guarantee the actor is loaded so GetSmartObjectComponentByRequestResult succeeds? In my scenario, this is happening on an actor that is loaded in (default OpenWorld world-partition level with all cells loaded) right next to me.

[crystal hatch]

Interesting. In principle, this could happen only for a smart object that has been added to the collection as part of collection building, and then gets streamed out at runtime. Can you confirm the actor is there but GetSmartObjectComponentByRequestResult fails? Which UE version are you using?

[normal mist]

Absolutely, will confirm. Booting up. Version 5.1 preview 2

It's taking me longer than expected to get results. When I remove the actor with the smart object from the level to validate that is the associated actor, my player's spawn gets despawned after a few seconds which prevents me from opening an in-game menu to check

okay got it. confirmed the actor

Pressed F8 after hitting play and I'm looking at the actor

[crystal hatch]

and you're sure the smart object you're handing is associated to this actor?

[normal mist]

I hit "Rebuild Colelction" on SmartObjectCollection and found 3 collection entries, one associating with each of the 3 "depot" actors, each of which has a smart object component

[crystal hatch]

Do you happen to hit a failed ensure when you end PIE?

[normal mist]

Querying USmartObjectSubsystem::FindSmartObjects(FSmartObjectRequest), I get three results, one for each of the bases. and for one of them it can't find the smart object component when running USmartObjectSubsystem::GetSmartObjectComponentByRequestResult

[normal mist]

will check now

[normal mist]

will this be in output log?

I did a ctrl + f in Output Log for "ensure" and didn't find any matches

message log looks fine, no errors or warnings

[crystal hatch]

you can use log LogSmartObject veryverbose console command to see all the logging SOs print out. Or use vislog, I believe it doesn't care about the verbosity setting of any given log category.

might contain hints.

[normal mist]

LogSmartObject: VeryVerbose: Registering BP_TerritoryControlBase_C /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.BP_TerritoryControlBase_C_UAID_E45E37FB55DACE4501_2092616428 using definition SmartObjectDefinition /Game/RoyalDivision/Features/Interactions/Capture/CaptureDefinition.CaptureDefinition.
LogSmartObject: VeryVerbose: BP_TerritoryControlBase_C_UAID_E45E37FB55DACE4501_2092616428 not added to collection since Main Collection is not set yet. Storing SOComponent instance for registration once a collection is set.
LogSmartObject: VeryVerbose: Registering BP_TerritoryControlBase_C /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.BP_TerritoryControlBase_C_UAID_E45E37FB55DACF4501_1154875607 using definition SmartObjectDefinition /Game/RoyalDivision/Features/Interactions/Capture/CaptureDefinition.CaptureDefinition.
LogSmartObject: VeryVerbose: BP_TerritoryControlBase_C_UAID_E45E37FB55DACF4501_1154875607 not added to collection since Main Collection is not set yet. Storing SOComponent instance for registration once a collection is set.
LogSmartObject: Main collection 'SmartObjectCollection_0' registered with 3 entries
LogSmartObject: VeryVerbose: Registering BP_TerritoryControlBase_C /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.BP_TerritoryControlBase_C_UAID_E45E37FB55DACE4501_2092616428 using definition SmartObjectDefinition /Game/RoyalDivision/Features/Interactions/Capture/CaptureDefinition.CaptureDefinition.
LogSmartObject: VeryVerbose: 'BP_TerritoryControlBase_C_UAID_E45E37FB55DACE4501_2092616428[2149769503]' already registered to collection 'SmartObjectCollection /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.SmartObjectCollection_0'
LogSmartObject: VeryVerbose: Registering BP_TerritoryControlBase_C /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.BP_TerritoryControlBase_C_UAID_E45E37FB55DACF4501_1154875607 using definition SmartObjectDefinition /Game/RoyalDivision/Features/Interactions/Capture/CaptureDefinition.CaptureDefinition.
LogSmartObject: VeryVerbose: 'BP_TerritoryControlBase_C_UAID_E45E37FB55DACF4501_1154875607[1782400616]' already registered to collection 'SmartObjectCollection /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.SmartObjectCollection_0'
LogSmartObject: Collection 'SmartObjectCollection_0' registration from USmartObjectSubsystem initialization - ESmartObjectCollectionRegistrationResult::Succeeded

This is only referencing 2 out of the 3, first hint

going through the log...

LogSmartObject: VeryVerbose: Registering BP_TerritoryControlBase_C /Memory/UEDPIE_0_MainGrid_L0_X-1_Y0_DL0.Level_Virtual:PersistentLevel.BP_TerritoryControlBase_C_UAID_E45E37FB55DA824701_2019018016 using definition SmartObjectDefinition /Game/RoyalDivision/Features/Interactions/Capture/CaptureDefinition.CaptureDefinition.
LogSmartObject: VeryVerbose: 'BP_TerritoryControlBase_C_UAID_E45E37FB55DA824701_2019018016[4231601148]' already registered to collection 'SmartObjectCollection /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.SmartObjectCollection_0'
LogSmartObject: Verbose: Adding SmartObject '1782400616' to runtime simulation.
LogSmartObject: Verbose: Adding SmartObject '2149769503' to runtime simulation.
LogSmartObject: Verbose: Adding SmartObject '4231601148' to runtime simulation.

LogSmartObject: Rebuilding collection 'SmartObjectCollection /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.SmartObjectCollection_0' from component list
LogSmartObject: Reseting collection 'SmartObjectCollection /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.SmartObjectCollection_0'
LogSmartObject: Verbose: Adding 'BP_TerritoryControlBase_C_UAID_E45E37FB55DACE4501_2092616428[2149769503]' to collection 'SmartObjectCollection /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.SmartObjectCollection_0'
LogSmartObject: Verbose: Adding 'BP_TerritoryControlBase_C_UAID_E45E37FB55DACF4501_1154875607[1782400616]' to collection 'SmartObjectCollection /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.SmartObjectCollection_0'
LogSmartObject: Verbose: Claim SUCCEEDED for handle Object:4231601148 Slot:3 User:1. Slot State is 'ESmartObjectSlotState::Claimed'
LogSmartObject: Verbose: Start using handle Object:4231601148 Slot:3 User:1
LogTemp: Error: SmartObjectOwner in UCaptureBehavior::Trigger does not extend team interface
LogSmartObject: Verbose: Released using handle Object:4231601148 Slot:3 User:1
LogSmartObject: Verbose: Claim SUCCEEDED for handle Object:4231601148 Slot:3 User:2. Slot State is 'ESmartObjectSlotState::Claimed'
LogSmartObject: Verbose: Start using handle Object:4231601148 Slot:3 User:2
LogTemp: Error: SmartObjectOwner in UCaptureBehavior::Trigger does not extend team interface
LogSmartObject: Verbose: Released using handle Object:4231601148 Slot:3 User:2
LogSmartObject: Verbose: Claim SUCCEEDED for handle Object:4231601148 Slot:3 User:3. Slot State is 'ESmartObjectSlotState::Claimed'
LogSmartObject: Verbose: Start using handle Object:4231601148 Slot:3 User:3
LogTemp: Error: SmartObjectOwner in UCaptureBehavior::Trigger does not extend team interface
LogSmartObject: Verbose: Released using handle Object:4231601148 Slot:3 User:3

Final bit:

LogSmartObject: VeryVerbose: Unregistering BP_TerritoryControlBase_C /Memory/UEDPIE_0_MainGrid_L0_X-1_Y0_DL0.Level_Virtual:PersistentLevel.BP_TerritoryControlBase_C_UAID_E45E37FB55DA824701_2019018016 using definition SmartObjectDefinition /Game/RoyalDivision/Features/Interactions/Capture/CaptureDefinition.CaptureDefinition.
LogSmartObject: Verbose: Removing SmartObject '2149769503' from runtime simulation.
LogSmartObject: Verbose: Removing SmartObject '1782400616' from runtime simulation.
LogSmartObject: Verbose: Removing SmartObject '4231601148' from runtime simulation.
LogSmartObject: 'SmartObjectCollection /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.SmartObjectCollection_0' ASmartObjectCollection::EndPlay - Succeeded
LogWorld: UWorld::CleanupWorld for Level_Virtual, bSessionEnded=true, bCleanupResources=true
MassEntity: New page: MassEntity
LogSmartObject: VeryVerbose: Unregistering BP_TerritoryControlBase_C /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.BP_TerritoryControlBase_C_UAID_E45E37FB55DACE4501_2092616428 using definition SmartObjectDefinition /Game/RoyalDivision/Features/Interactions/Capture/CaptureDefinition.CaptureDefinition.
LogSmartObject: VeryVerbose: Unregistering BP_TerritoryControlBase_C /Game/RoyalDivision/World/Levels/Virtual/UEDPIE_0_Level_Virtual.Level_Virtual:PersistentLevel.BP_TerritoryControlBase_C_UAID_E45E37FB55DACF4501_1154875607 using definition SmartObjectDefinition /Game/RoyalDivision/Features/Interactions/Capture/CaptureDefinition.CaptureDefinition.

[crystal hatch]

something does seem off, but it's hard to tell what exactly without stepping through code. Sorry.

[normal mist]

no worries, I really appreciate you taking time out of your day to chat

thanks for all of your work on this system & Mass!

Are you aware of any literature online detailing how to properly set up smart objects in World Partition such that they are registered properly into the subsystem & collection? I am just hitting Rebuild Collection on the SmartObjectCollection actor in the level. I wouldn't be surprised if I was skipping a step

I am aware of the commandlet

Not sure how I would be regenerating the list at runtime though to account for a smart object attached to a moving actor

[crystal hatch]

Are you aware of any literature online detailing how to properly set up smart objects in World Partition
Like I said, it's not really supported, so.. no 🙂

[crystal hatch]

Right now moving smart objects are not supported. This will change, but it's another limitation of current design.

[normal mist]

Ill take a look at how Lyra handles their capturable control points since that is effectively what I'm implementing

I'm building out a debug menu which shows what interactions are near a player, and I am currently using Smart Objects + UGameplayBehavior as the foundation for the interaction system

[crystal hatch]

we are working on a "gameplay interactions" system that's an AI-independent extension of the SmartObjects concept, but it's very early in its development so don't hold your breath 🙂

[normal mist]

Awesome! I feel so fortunate to be able to look through the engine and sample project's source code to see how folks are going about these implementations. Not being in a big company anymore, I miss that ability to look through company wide repos to see how folks who are really knowledgeable about a specific use-case do their implementation. Using Unreal Engine gives me a bit of that with the source code being available.

I guess there's nothing stopping me from making multiple drivers for my implementation system (i.e. one using Smart Objects, another using a very lite home-grown system that gets me as far as I need to go until Smart Objects is feature complete)

I already have my own UMyCoolGameInteractionSubsystem which provides the standardized interaction API ontop of whatever the underlying foundation is, whether that be smart objects or something else

[crystal hatch]

UnrealEngine's full source being publicly available is a huge boon for anyone interested in how stuff is done in the biz! It's also a real picture of large code base - deprecations, legacy stuff, stupid mistakes (it's just a matter of scale) and so forth. I wish it was the case when I started in the industry and had no idea how stuff really works under the hood.

[normal mist]

It's been fun seeing the recent shift towards FSomeFeatureHandle and it being interacted with through USomeFeatureSubsystem in various new UE features

[slow sphinx]

Yeah the navmesh is there, I should have posted the pictures...the event tick on the Player Controller is working and as far as I can there is nothing wrong with the Character BP.

[crystal hatch]

vislog, what does it say?

[slow sphinx]

the term escapes me sorry, I thought you were referring to the visual aid for previewing the navmesh. I'm opening the project now

I was unaware of the vislog @crystal hatch Thanks for that

not sure what it means though 'Start point not on navmesh'?

[crystal hatch]

it should draw the relevant locations in the viewport. It means the system was unable to project the point to navmesh. So either no navmesh in that area, or the starting point was too high above the navmesh.

[slow sphinx]

This is what I'm seeing for the navmesh...that's really confusing

I was standing in the center of the picture when it was doing that

@crystal hatch I may have messes with the Navmesh settings for the project...is there an easy way to reset those settings?

[crystal hatch]

the easiest for me is going to DefaultEngine.ini and removing all the relevant bits (the classes I want to reset).

[slow sphinx]

Ah alright I'll try that, I guess I could also open another project and just export from their and import to this one...you would think there would be a reset to default option

[crystal hatch]

there might be, I just never used it 😄

[slow sphinx]

Yeah I reset the settings I think but it didn't seem to change anything...unless that's a bug...anyway that didn't work

[slow sphinx]

Fixed it

The Navmesh bugged out

that's all it was....yet another UE5.0 bug

I have no clue why it seems I keep coming across every UE5.0 bug in existence

if you change anything on the Navmesh settings in UE5 it seems likely it will bug out the Recaster
that I am assuming is the cause of the bug

[crystal hatch]

if that's the case then it's an unknown bug. Have you tried debug drawing navmesh at runtime? Is your game Client-Server or standalone?

[slow sphinx]

It's standalone, I didn't try debug drawing it at runtime...but I'm fairly sure it was a bug between the RecastNavMesh actor and the Project settings....this shouldn't surprise anyone considering the old one let you change the settings from the actual actor but doing so never worked.

They did change the RecastNavmesh actor in 5.0

Everything I have noticed so far that they have changed has one or more bug so far...I've literally not found one thing they changed that wasn't ate up with bugs

Can't wait for 5.1

I'll look into that debug drawing though at runtime, I've done it once but don't remember how. Thanks for introducing me to the Visual Logger.

[crystal hatch]

this shouldn't surprise anyone considering the old one let you change the settings from the actual actor but doing so never worked.
This is a surprise to me since I've been changing local navmesh actor's properties for years now and it always works - only the agent-related properties are off limits, by design (and thus should be hidden in properties, I agree).

[crystal hatch]

No, not really. Re bugs - I'd love to hear some details.

[slow sphinx]

They removed the Navmesh options from it

Nevermind

I thought they did

lol

I never touch it anymore and I thought I checked and didn't see them. I know that in UE 4 changing it from there never worked so I must have assumed they were removed finally

It would but restarting would reset them

Maybe it's only meant for testing or perhaps the link between that and the project settings were never setup? Idk

[crystal hatch]

Not true. The navmesh instance gets discarded only if it doesn't match declared SupportedAgents and these check only agent-related properties, so you can change other navmesh generation properties as much as you want.

[slow sphinx]

Maybe I'm mistaken but I clearly remember changing them from the RecastNavmesh actor and it didn't work in UE4...I even remember spending days trying to figure out how to do it properly until I came across a video showing how to get around that by setting it from the Project Settings

[crystal hatch]

whenever anyone reported that as a but it turned out they were messing around with agent properties 🙂

[slow sphinx]

but if you can't set that from there then why have it?

[crystal hatch]

that's why I said that these specific ones should be hidden in instance's properties display.

[slow sphinx]

when I was first learning about nav meshes that really threw me off...like I said I spent days trying to figure out how to do it

[slow sphinx]

Oh right

I haven't touched them that often since then because I've been exploring other areas but I guess I'll have to refamiliarize myself again

Thanks for the talk and the info, appreciate it

[crystal hatch]

np

[vast relic]

Who here is the navmesh guru?

I'm working on a vision and influence mapping system, I wonder if I should piggyback on the navmesh instead of doing my own thing in parallel

[meager galleon]

Quick question, I'm implementing a drone companion for my character and want it to follow the character (among other things), would it be smart to add a Key to my blackboard with the player's location and constantly update the value in the AIControllers tick function?

Something tells me there must be a better way.

Nevermind, just answered my own question. I just update the location in a task of course.. Overthinking.

[celest python]

You should prefer having an AActor key and use MoveTo node by managing the state/flow of the behavior in BTs

[crystal hatch]

I suggest using EQS for drone's positioning. In Bulletstorm we used EQS's predecessor to pick nice spots close to the edge of the screen so that the sidekicks were visible, but not obstructing the view.

[lethal helm]

is there a way to enable debugging or logging for navlinks?

I know the navlink line fails to draw if the navlink is invalid for some reason but I haven't figured out how to get it to tell me why the navlink is invalid

[celest python]

I didnt use navlinks so far but check visual logger

---

is there any thread safe way to project a point to navigation?

without accessing any type of UObject

[celest python]

I did it

it just needs more love to stop creating weird curves like this but if I ignore this result looks like 1:1 to Death Stranding's method

[lethal helm]

I hate smart navlinks so much

I'm somewhat tempted to just add a sphere collision to the non-smart navlinks so I can get nav link started event without the bugginess

[uneven cloud]

At that point it would be cheaper to use smart links. If you use C++, it's not buggy at all to use point links and nav areas.

[lethal helm]

I'm using c++

sometimes the smart links work sometimes they don't. sometimes they work and appear visually while debugging navigation sometimes they don't

sometimes they don't appear visually yet still work

Works perfectly

FVector SpawnLocation = FVector(0, 0, 0);
FRotator SpawnRotator = FRotator();
ANavLinkProxy* NavLink = GetWorld()->SpawnActor< ANavLinkProxy>(FallNavLink, SpawnLocation, SpawnRotator, SpawnParams);
NavLink->PointLinks.Empty();
FNavigationLink PointLink = FNavigationLink(GridPointArray[i], FoundPoint);
PointLink.Direction = ENavLinkDirection::LeftToRight;
NavLink->PointLinks.Add(PointLink);```

Only some of these links work and they aren't visualized by nav debugging

FVector SpawnLocation = FVector(0, 0, 0);
FRotator SpawnRotator = FRotator();
ANavLinkProxy* NavLink = GetWorld()->SpawnActor< ANavLinkProxy>(FallNavLink, SpawnLocation, SpawnRotator, SpawnParams);
NavLink->PointLinks.Empty();
NavLink->GetSmartLinkComp()->SetLinkData(GridPointArray[i], FoundPoint, ENavLinkDirection::LeftToRight);```

also works perfectly

FVector SpawnLocation = FVector(0, 0, 0);
FRotator SpawnRotator = FRotator();
ANavLink_Climb* NavLink = GetWorld()->SpawnActor< ANavLink_Climb>(ClimbNavLink, SpawnLocation, SpawnRotator, SpawnParams);
NavLink->GetSmartLinkComp()->SetLinkData(OutPoint, Start, ENavLinkDirection::LeftToRight);
NavLink->PointLinks.Empty();
NavLink->ClimbPoints = ClimbPoints;```

[lyric flint]

is there any way I can have my AI be stuck to a single axis via the blueprint chart? I've tried a lot but the AI doesn't respond to it. I've tried constrained to plane and blocking off the other axis with the volume modifier.

[lethal helm]

what do you mean? Navigation is constrained to the horizontal plane by default

[lyric flint]

like I want to constrict it to the Y-axis

so my AI can just move up and down only

[lethal helm]

oh you will want to make a custom ai movement system then

[lyric flint]

is that possible in BP?

[lethal helm]

no

[lyric flint]

oh

then I'm kinda screwed

oh geez

I appreciate the help

[harsh storm]

It is possible in BP.

Just a lot more laborious 😅

[lyric flint]

you say

it's possible

where can I start?

[harsh storm]

Just create an actor component

Then start doing what you need to.

You will be starting from absolute scratch mind you.

So you won't have nav support out of the box.

You can do it in BP, but it 100% is not worth it.

Just pick up a little bit of C++ and then create what you need.

[lyric flint]

alright

[harsh storm]

All that said - I'm pretty sure the movement components can already restrict movement 🤔

There are settings here in the CharacterMovementComponent for example

It's also in the FloatingPawnMovement component

[lyric flint]

I have that enabled on the AI but he will just move a bit but then sit there

it's weird

[crystal hatch]

How do you make your AI move? If you’re using pathfinding at all then there’s your problem - the path probably isn’t restricted to your plane, so character’s movement doesn’t match path following component’s expectations, thus path following failure, thus stopping.
Your AI should be able to move if you just use “direct” movement commands (i.e. without using navigation).

[winged zinc]

Poll on visual node based no code ai development (as seen in unreal and unity for games) https://www.reddit.com/r/artificial/comments/ypye9v/how_many_of_you_would_use_a_drag_and_drop_visual/

[lyric flint]

I'm using the basic Navmesh. It seems to me you are right because that could be the reason. Do you mean just go into the BP itself and just write down movement commands? I just want the AI to roam about the one plane without moving one unit into the other and still doing the whole AI tree commands.

[crystal hatch]

I believe you're mixing two concepts:

1. picking where to move and
2. moving there

Limiting movement plane will only influence point 2. But point 1 is where you need to start - if you pick points outside of the movement plain the move will always fail.

If you pick points in the plane you care about then you can just use Move Directly Toward BT task.

[lyric flint]

hmmm, so I need to pick the axis of where to move before I start to move there?

[crystal hatch]

no, you need to pick move goal location that's on your desired movement plane.

[lyric flint]

oooooh

I think I understand

[ocean wren]

looks good Eren!

[ocean wren]

Yeah, you mean like Azure AutoML? 🙂

Hmm, whats the preferred way of capturing mic audio in unreal do you think? there's an onlinesubsystem for voip, but I dont need the online bit 🙂

[celest python]

I think there was a new AudioCapture component in 5.0, maybe it can help?

[ocean wren]

Yeah, maybe thats the one.. I used that before.. but I don't really want to use a component for it.. cos I'm after doing some editor tools using it too

I guess I could look at how it works and just see if I can make it a subsystem

First, I'll have a go with FVoiceModule and see if that works

[winged zinc]

But with ability to make minute changes and export to python

[ocean wren]

There are a bunch of apps like that

I mean its a fair idea, just a lot of competition

[median arch]

anyone able to help me debug my Perception Component? It suddenly seems to have stopped firing OnPercptionUpdated events and I don't know why.

Controller constructor:

    AIPerceptionComponent = CreateDefaultSubobject<UAIPerceptionComponent>("AIPerception");

    UAISenseConfig_Sight* SightSenseConfig = CreateDefaultSubobject<UAISenseConfig_Sight>("AISense_Sight");
    SightSenseConfig->PeripheralVisionAngleDegrees = 360.f;
    SightSenseConfig->DetectionByAffiliation.bDetectEnemies = true;
    SightSenseConfig->DetectionByAffiliation.bDetectNeutrals = true;
    SightSenseConfig->DetectionByAffiliation.bDetectFriendlies = true;
    SightSenseConfig->SightRadius = 5000.0f;
    SightSenseConfig->LoseSightRadius = 7000.0f;

    AIPerceptionComponent->ConfigureSense(*SightSenseConfig);
    AIPerceptionComponent->SetDominantSense(*SightSenseConfig->GetSenseImplementation());

Controller PostInitializeComponents:

void AEFAIControllerCPP::PostInitializeComponents()
{
    Super::PostInitializeComponents();

    AIPerceptionComponent->OnPerceptionUpdated.AddUniqueDynamic(this, &AEFAIControllerCPP::OnPerceptionUpdated);
}

And the callback is now just

void AEFAIControllerCPP::OnPerceptionUpdated(const TArray<AActor*>& Actors)
{
    for (AActor* Actor : Actors)
    {
        UE_LOG(LogTemp, Log, TEXT("Observed %s"), *Actor->GetName());

    }
}```

It's definitely marked as a UFUNCTION:

UFUNCTION()
void OnPerceptionUpdated(const TArray<AActor*>& Actors);

Moreover I'm certain this used to work and have no idea why it has suddenly stopped...

It's like UAIPerceptionComponent::ProcessStimuli is never called... hm

[slow shadow]

I'm trying to make a, hopefully, basic behaviour tree. I followed a few guides but my "rotate to face BB entry" is getting stuck. Any tips for a guide?

[median arch]

whaaat, Senses Config is an empty array...? ->ConfigureSense(*SenseConfig) didn't throw any errors... even moving to beginplay keeps the array empty at runtime... ha

If I click the + on the Senses Config array, and then choose AI Sight config, it populates the sense with all the values I set in the constructor... 🤷‍♂️

(I changed 7000 to 8000 to double check that)

Wow... After searching for Sense config I found this: #cpp message

And so I tried renaming the components in CreateDefaultSubobject.... and it worked

[polar crown]

Hi guys, anyone have any idea why the area underneath my AI model isnt working with navigation? he just gets stuck in this square when trying to move. i have deleted each aspect of the model and nothing seems to change. thankyou in advance

[brittle lynx]

Disable "can ever affect navigation" on the character

and anything attached to the character

[polar crown]

hi that is selected because i had the floor selected. it is off on the character

[brittle lynx]

is there a weapon etc. that the character is holding? That could be blocking navigation as well.

[polar crown]

i have gone down the list of what is included in the character, sword, shield etc and they all have that option turned off... strange

i have a feeling its the mesh but i cant see any options that would effect this

sorted

[meager galleon]

Hmm, I have a very simpel EQS query that just finds the character on a test map. Thing is, the grid points does not seem to work when the AI is in the air. Anyone know if there is a solution for this?

[celest python]

increase projection extent

[meager galleon]

Just started with EQS today and is quite new to game dev as a whole, where can I find projection extent?

Nevermind, it was the Pathfinding grid I tried with, that one seems to only work on groundlevel.

[crystal hatch]

Make sure your character's bUseControllerRotationYaw is true

[crystal hatch]

To be precise, it only works on navmesh.

[shrewd pumice]

Heyo, using BTs and AI Perception, is hearing sense the best way to implement a detection radius even for stationary actors?

Example: A patrolling ai wanders past a point of interest (a dead body) and stops to investigate the area.

Should I put a sound stimuli on the actor, or should I just use a collision box with event overlaps?

[shrewd pumice]

Hmm, would it be a problem if these things were always making sound? Perhaps a blueprint interface with box collision would be the smarter way...

[keen crow]

if you want to detect dead bodies why don't you put a sight stimuli component on dead actors?

[shrewd pumice]

Ah, well I'm making a game where the primary antagonists are ghosts. While they randomly patrol they should be drawn to dead body actors, and if anything that's living enters a certain proximity radius they will aggro. What do you think would be a good way for them to detect things in such a scenario?

"Sensing life force" as the top priority sense, followed by sight.

I'm just not sure if using the sound stimuli would be the best way to implement proximity sensing.

[keen crow]

well if they should be drawn to dead bodies during patrol and then attack player on sight I'd suggest to

1. attract ghosts to dead bodies either by EQS queries OR put smart object components on dead bodies and utilize smart objects spatial lookup system
2. do nothing with players since by default pawns have sight stimuli component implicitly added to them unless you disabled it manually or not using pawns for players at all

[shrewd pumice]

Thanks! I'll look into your suggestions and we'll see if I can work it out.

[severe tangle]

I'm using EQS, where the target location is picked from 25% best matches, which resolves in AI looking more believable.

The issue is with the path itself, where AI always tries to take the shortest route, meaning cutting sharp corners and hugging walls.

Is there a way how I can achieve more believable pathing with EQS system?

[meager galleon]

I'm trying to use MoveDirectlyTowards in my BT for my flying AI ACharacter. I've eneabled CanFly and set FlySpeed, but the AI won't move. I've also made sure TargetLocation has been set correctly. Is there something I'm missing here?

I've also tried with Gravity off for both StaticMesh and Capsule.

[crystal hatch]

check if there's anything in vislog, might contain some clues

[meager galleon]

Never used that tool, don't really know what I'm looking for to be honest

[crystal hatch]

the log says the movement task is invalid. I'd check the output log for errors and or warnings. Also check the vislog for the same - easy to find since errors and warnings are marked on the timeline.
Some more details would be helpful, like which movement component are you using?

[meager galleon]

I solved it, I had added a StaticMesh to my character which I used just for a dummy cone, turns out that does not move, but my SkeletalMesh actually did.

Thank you though.

[slow shadow]

That helped a lot, thanks mate

For some reason my BT goes to Idle after one hit from combat, shouldn't it refresh combat selector as long as the target is in sight?

[crystal hatch]

it does exactly what you told it 🙂 So double check your BTs logic (i.e. structure). VisLog helps a lot with that, since it can tell you step by step what happened and what was the state of the blackboard at that time (with a caveat that BB state is captured along with the very first log line in a given frame).

[slow shadow]

But I have it on "aborts lower prio" while blackboard/AI has a PlayerTarget, it doesn't make sense.

Do note that I'm very fresh to BTs. What is this "VisLog"? An external plugin?

[crystal hatch]

But I have it on "aborts lower prio" while blackboard/AI has a PlayerTarget, it doesn't make sense.
But if the condition fails it will still go to Idle. It will also go to Idle if all the options under Combat fail.
VisLog is UE4/5's native AI and Gameplay debugging tool. My favorite way of summoning it is running vislog console command. There's a way to open it form editor as well, but I never remember how.

[weary holly]

has anyone ever created new custom flow nodes for the behavior tree?

[celest python]

if you mean composites, i tried

its pretty easy to do unless you want multiple outputs

i lost my sanity after trying to understand how simple parallel implements two output pins

[ocean wren]

That implies you had some sanity to begin with

[opaque whale]

Be careful making composites, it can get violent

[weary holly]

👀

[silent drift]

How can i set a blackboard value when AI sees a player and unset it when AI doesnt see the player anymore?

[lethal helm]

so weird scenario

I've got a different nav meshes at different altitudes

and the pathfinding seems to think a path teleporting from the lower nav mesh to the upper nav mesh from points A to B is legal

[uneven cloud]

If you are using the perception component, you can bind to on perception updated. You can also use the EQS if you have more than 1 possible target.

[silent drift]

If i use the perception updated, where should i unset the blackboard value to tell the behavior tree that AI doesnt see the player anymore

[uneven cloud]

You can bind to OnTargetUpdated when you first see it. However on perception updated should also be called when sight is lost.

[silent drift]

I did use target updated before, but i do not know where to put in the unset blackboard value

[uneven cloud]

You can call Clear blackboard if successfully sensed is false

[silent drift]

when successfully sensed false?

Okay! I will try that. Thank you

[shrewd pumice]

Heyo, How do I make sure my ai controller is not sensing noises that it is making itself? I mean that in an instance specific way. The problem occurs at that third branch node.

In this an example, a mob of ghosts is making noise that will aggro other ghosts they pass while chasing the player.

[severe tangle]

When AI picks a path with EQS, can I somehow get the nodes in this path and make them more expensive to travel, so that other AIs will have lower chance to pick the same/similar path?

[uneven cloud]

Not without a lot of C++

[uneven cloud]

You want to check if the sensed actor is not the pawn. Self is likely the controller, especially based on your comment.

[severe tangle]

I see, well then randomly spawning more expensive nav areas at AI location when he is moving to the target will have to do 😅 . Not sure about other approach to make AI's pathing more believable and not always the same/the shortest path to the target.

[uneven cloud]

That's a really bad idea. It's better to use avoidance and adjust the destination.

[severe tangle]

I was using RVOAvoidance and loved it, the issue was that its not really usable in multiplayer game, because clients were receiving huge amount of network corrections

[charred lava]

For AI's Hearing Sense, do AI's standing closer to the Noise Event Location process it before the ones further away?
I want to make sure that only the closest AI will handle the Noise Event.

I could wait one frame and check every AI that registered that noise and then compare the distance and let the closest react, but if there is an easier way I take it ^^

👀```c
class AIMODULE_API UAISense_Hearing : public UAISense
{
/** Defaults to 0 to have instant notification. Setting to > 0 will result in delaying
* when AI hears the sound based on the distance from the source */
UPROPERTY(config)
float SpeedOfSoundSq;

[crystal hatch]

there is indeed a distance-based delay for hearing stumuli, but it will be eventually received by all the listeners within effective range. You'll need to add some ticketing system to have only the first one react to the stimulus.

[charred lava]

I'm on it 👍 Thanks MieszkoZ

[keen crow]

why do I no longer see EQS items scores on the spheres when I enable AI visual debugging (apostrophe). It was working in UE4.26 but it seems that by default it doesnt work in UE5.03. I can still see the scores in the query details table by pressing '/' but I'd really like to see scores on the items themselves

[hoary hawk]

Good morning everyone ! I have a project where I have 50-100 AI moving around in the scene, I did all the optimization I could possibly think of, except one and it seems to be the biggest memory eater : characterMovement. I am using the default component from unreal, I know I already read that it does take a lot of ressources. I was looking for other alternatives but could not find ressources on doing specifically this for AI, and everything it could include like getting NAV path from navmesh etc. Where could I start? Do you know any ressources / documentation to implement this ?
Thank you !!

[manic karma]

Have you used the profiler/insights to verify that character movement is your issue

[hoary hawk]

Only tests I did was running the stat character command in the game, Nd disabling on command the movement of on AI at runtime. When I do the char movement goes to normal ( 1 call per AI), but when I activate it back it goes crazy for all AIs

[storm zephyr]

hello, by enabling the "Do Fully Async Nav data gathering" would that also update the navmesh async at runtime or that is just for the building of the navmesh at design time?

[severe tangle]

With Move To Location I can set Filter Class dynamically via variable, but when using Behavior Tree's Move To node, I can only set this filter manually, I don't see an option to use a variable, that I can later change, so that AI is using different nav filters from time to time.

[celest python]

you can either inherit from UBTTask_MoveTo in c++ and listen for blackboard keys and update your filter class when its changed and recall MoveTo() or create a BP and use MoveTo node task and work with delegates + service

[uneven cloud]

I've never seen this. How many AIs do you have bumping into each other? Regenerating nav mesh for a moving character is going to be incredibly expensive. It also invalidates paths, so you'll be repathing a lot.

[severe tangle]

Could you please elaborate more with the latter option? I'm not sure if I understand.

[severe tangle]

Well basically I have a Character that can be both player and an AI, and inside CMC I set up to use RVOAvoidance, on server I loved how it worked, exactly what I needed, but then when I tried using it on client, I couldn't even move forward without getting rubber banded right back, like the client doesn't even care about RVOAvoidance radius and server is sending him back because he is colliding with his own radius or I have no idea. Then each time client hit other AI, he once again completely ignored this RVOAvoidance radius, going right through the AI, but then server started heavily correcting him because he should've been "pushed" to the side.

[celest python]

so MoveTo node in BPs and BTTask version of MoveTo is almost identical, only difference is in BT graph you are not able to edit the non blackboad values to change the state of the current move to task. If you create a new BTTask and use BP MoveTo node inside of it, you can use delegates to notify your BTTask and re-trigger the MoveTo with new params you provided

[severe tangle]

Oo now I understand, okey if MoveTo inside BP is almost identical to BTTTask version, I will just use the BP's one, thanks a lot!

[celest python]

yeah only difference is one of them is able to listen BB values other one is not

[uneven cloud]

Because the server is the authority for server owned actors. Does RVO not work for the server and correct on the clients? You might get better results with crowd following.

[severe tangle]

RVO works on the server, doesn't seem to be working on clients, that's why they're getting spammed with corrections.

Crowd Following is one of the new UE5 features? Trying to google it but there doesn't seem to be much info

[uneven cloud]

I mean, having RVO on the server and not on the client. Letting the server control the movement.

[uneven cloud]

No. It's been around for a long time

[celest python]

how RVO even works in client anyway?

client's laggy movement can cause some distortions in avoidance and it might look weird in client

[severe tangle]

Ooh it's the Detour Crowd? Because if so, I was playing with it, but the jittery movement was really bad

[severe tangle]

It was pretty much unplayable from the client's side

[celest python]

with how much ping you tested?

[severe tangle]

I tested it on both decent hardware and net connection all the way to the very poor hardware and garbage internet connection, ping around 500, at that point it was a circus 😅

[celest python]

500 is "huge"

its only to test critical features

it takes half of a second to arrive the information to client

ofc it will look bad 😄

some games use lockstep networking which UE doesnt support, when client goes out of sync it continues to simulate the world with the last info it received

in client-server architecture you have different advantages and disadvantages

[uneven cloud]

A ping of 500, all movement is going to look terrible.

[severe tangle]

well with CMC it's actually not bad at all, only the RVO was doing ruckus

[sacred shale]

does anyone know an elegant solution to AI listening for both pawns and distant sounds? I've been using the pawn sensing component on my AI as well as a pawn noise emitter on my main character, but it feels like it's only good for pawns with noise emitters and not regular sounds. been trying to look into something consistent but haven't had any luck

ideally the same component should be listening for both pawn footsteps/sounds as well as distant sounds in 3d space

[lethal helm]

I've got a different nav meshes at different altitudes
and the pathfinding seems to think a path teleporting from the lower nav mesh to the upper nav mesh from points A to B is legal
What could be causing this?

[haughty coral]

Happened to me before with nav projection extent if dest was on the mesh above the lower mesh. I think default extent is in project settings but you can override it at many places. Iirc also projection in character movement comp was causing the char to incorrectly switch between these. But if this us happening mid path (not near path start/end) then I am not sure. Maybe check vislog / gameplay debugger.

[lethal helm]

which setting is default extent?

[haughty coral]

Default Query Extent

try to lower Z

and delete your RecastNavMesh actor and rebuild navigation again

[lethal helm]

didn't help

what would I even be looking for in the viz log

[lethal helm]

well the current solutino is just making those areas not part of the navmesh

which sucks but there doesn't seem to be any alternatives

[lethal helm]

also this comment makes it seem like query extent is the agent size

[uneven cloud]

The pawn sensing component is old, recommend switching to the perception component where anything can report a sound.

[uneven cloud]

Is that the actual path or do you allow partial paths? Check the vis log to actually see the path it is finding.

[lethal helm]

nooooo

that was it

I hate my predecessor sometimes. why was that even turned on

[lethal helm]

thanks. that saved me some of my remaining sanity

[uneven cloud]

You're welcome. Partial paths is turned on by default and always causes odd behavior.

[lethal helm]

hmm still happens but much less often

[rich veldt]

Hi, I'm trying to brainstorm a solution. I have points an AI can run to using EQS (cover points), and I'm trying to filter out any points that would result in the AI running past a context (their target). I want them to pick the furthest cover point from the target which is fine, but sometimes that results in them running right at the target which isn't really "running to hide" behavior.

Just drew up an example: it's scoring the green box lower because it's not as far from the target as the red box, but the red box should be disqualified because it means running past the target. Not sure how I could check for that reliably.

Maybe just disqualify any that are in the direction of the context +- 90 degrees? Wouldn't be totally reliable but would at least reduce the amount of times this happens. Would still like a solution that essentially checks if the pathfinder path to the item goes "near" the context.

[keen crow]

you could try either

1. create your custom test that takes the target player as the context and filters out items in angle range [- X degrees + PlayerRotation; + X degrees + PlayerRotation]
2. add an extra distance test but this time between an item and the player and make a negative score for this test so that it drags down the score of items which are closer to the player

or try using both of them

[rich veldt]

Correct me if I'm misunderstanding you here, but 1. just sounds like my suggestion under the picture? and 2. sounds like what i'm already doing (closer to target = less score, farther from target = higher score)

2. doesn't work on its own because if the farthest point from the target is way behind them, then the AI would have to run past the target to get to it and that doesn't make sense.

adding 1. in conjunction with 2. would filter out those points^, and ensure the AI always runs away from the target, but it would prevent the AI from considering cover points which are between them and the target. Here's an example of a situation where using an angle to the target would disqualify totally valid cover points.

"most viable" is no longer allowed because it's within the angle which is meant to disqualify that "furthest point".

If there were cover points behind the AI, to the left of the image, then he'd go there and that's fine. If there weren't, then he just wouldn't take cover. So it's not a perfect solution but this is the only one I have.

Maybe only disqualify items in that V which are further from the AI than the target is?

[keen crow]

is the black circle here is NPC? if so, are those green lines are angle filtering? are you filtering by the npc's field of view?

[rich veldt]

Yes the black circle would be the NPC. I'd be filtering by direction from NPC to Target, with an FOV of like, nearly 180 degrees (-90 to +90)

[keen crow]

but why don't you want to filter out items that are in the field of view of the player if you want to make an NPC avoid players FOV?

oh right

because some item can be out of the players FOV but still an npc would have to run through the player

[rich veldt]

That's not what I'm trying to do.
I'm trying to make the NPC run somewhere "deeper" into the location and hide in ambush. It gets all possible ambush points in a radius around itself, then determines the best place to hide.
The checks are:

• Cover point is not in use
• Cover point would not be "exposed" to the target should they come towards it (no parallel covers) : i.e if the cover point is opposing +X, the target must be on the +X side of it
• Cannot be in line of sight to the target
• Prefer the cover point which has the furthest path from the target
• Is a reachable point via pathfinding for the NPC

[rich veldt]

So yes, the furthest ambush point from the target (which is also an opposing-facing cover point) could be "behind" the target - requiring the AI to run past the target in order to go and hide.

[keen crow]

you've got me interested. I'll try to draft something on an empty project out of my own curiosity

besides, I might be doing my own cover system for AI some time soon

[rich veldt]

I'm thinking about making a test which generates a nav path to the item, then loops through the points in the path and checks if the points "pass" the target relative to the NPC (or get closer in distance to the target).

[keen crow]

i believe the answer might be easier

[rich veldt]

I mean a sphere trace toward the item could be OK. If it hits the context/target it's a bad item. Would be iffy in weirdly shaped corridors though and adds cost.

Ohh...maybe I could generate a path to the item, get the first point in the path that's further than the target is, and check if the direction from the NPC to that next point is within like 180 degrees of the direction from the NPC to the target? If so, then disqualify it because it's requesting the NPC runs "toward" the target **and **"farther" than it - which would sort of indicate "passing" it wouldnt it?

Complicated but seems not terrible. NPC breaks out of "run and hide" sequence if they get within combat range anyway so player would be none the wiser that it was a failed sequence. I'm open to an easier solution though 😛

[keen crow]

How and where are you going to generate path from npc to best cover point in eqs items?

[rich veldt]

idk id just code a new test in c++ to do it

i have a custom pathfinding component on all npcs that i can request to make a temporary path very easily

[keen crow]

hmm, have you thought of doing that with 2 eqs queries? one to find the best spot, second is to find the best proxy point to reach it avoiding target's FoV?

perhaps the second query might be fired in a service while running BTTask_MoveTo cover

because the player might be looking around and stuff

[rich veldt]

well it's not related to the fov of the target/what the target can see.

the target could be turned around/not looking at the AI and they'll still think "oh shit someone's here i'd better run deeper into the tunnels and hide around a corner"

im gonna try my method i mentioned above later

[keen crow]

well if you don't consider player's FoV while an npc is getting to the cover player might change its position and rotation and that cover won't longer be viable

wouldn't it?

[rich veldt]

Its supposed to work like hide n go seek/tag. You are spotted, they run off to hide, if you catch up to them before they get to their hiding position yes they will cancel and face you. if you dont, they will try to surprise you.

I have a behavior tree node checking self visibility against the target's pos and fov yes, thats how you can force a cancel (by chasing them).

But the method of choosing isnt related to if the target can or cannot see it. It's filtered by trace from target's location to the cover point- so even if they arent facing a cover point it will disqualify points that are "in the open" so to speak

Oh I see what you mean. There's also already a Behavior Tree node that checks if the currently occupied cover point is visible by a target which is currently visible by the NPC, and it will leave said position and either engage or move accordingly or cancel moving to it or something. Yeah.

but that's sorta beyond what im talking about which is simply just considering the initial cover position and how to not move PAST the target

rather, to not consider at all, any cover positions which are beyond the target

[keen crow]

hmm, considering your gameplay requirements one other thing I'm thinking right now is to make an eqs test that should be placed final and it should test each remaining eqs points from best to worst in the following manner: cast a cone-like shape from the player (like player's fov) and then do a shape sweep by object channel from the querier to that item and if it overlaps player's fov cone then this point is invalid. But I feel like it's only going to work in a simple and flat level geometry 🤔

and I have no idea how to cast a single-use shape of such form like that from the player. perhaps make it a collision component attached to the player 🤔

[rich veldt]

you could do all that with some simple vector math/dot product

then acos degree and check against the target's fov (is less than?)

but that's not really related to the issue because like i said all points that are in line of sight to the target (whether or not they're actually looking towards it) are disqualified anyway. the player looking toward the point is irrelevant since it's already confirmed to not be within line of sight.

and if you dont do that line trace, and just do what you said, AI will avoid all points the player is facing even if he can't see them. so it's a moot check regardless

[keen crow]

but what about points that are out of players fov but getting to that point would require the npc to enter player's fov? those are not discarded are they? this is where that test I thought of comes into play. like ok you've got a good point outside of players fov but would getting to it require you to cross players fov?

[rich veldt]

i dont care about that. it's a horror game so seeing something run through your LoS that you didn't even notice before as it disappears running down a hallway is cool

[keen crow]

then it seems I failed to understand what you actually need 🙈

[rich veldt]

yeah i know lol

i thought i was helping you lol

since i already solved my problem when i said i will get the intended path to the "winning" cover position, and ask if its points indicate movement toward and beyond the target. if so then disqualify it and go to the next "winning" position and do the same test.

[silk fulcrum]

Does anyone have any basic wisdom on using the new state tree as the basis for my AI vs of I should stick with behavior trees?

[ocean wren]

Probably depends how brave you are 🙂

Its still being worked on afaik, so you know..

[rich veldt]

just a btw but this worked. just finished programming it. ai won't go to a cover point (only test the winning ones!) if the path to get there has 2 points one after another, the first being closer to the AI than the target is, the next being further than the target, and the direction from the closer point to the further point is within +-90 degrees of the direction from the closer point to the target. indicating that we're likely "passing" the target at some point during that path, which is no bueno.

[uneven cloud]

What you want is the dot product with a filter. Dot of target to querier and target to item. I'd recommend turning off all other tests and using the visual logger to get the correct settings. Significantly cheaper than checking path points.

[cinder garnet]

hey can i get someone to help set up my roaming Ai, finally got time to do it

[celest python]

https://dontasktoask.com/

[cinder garnet]

bruh

[zealous sierra]

Does anybody know of any Unreal games that ship with the ability for players to make their own AI?

[celest python]

best thing comes to my mind that is any UE game has modding support?

[zealous sierra]

That is what I feared 😄
I couldn't find anything that let players use a scripting language or something in a way that didn't involve downloading a whole dveeloper kit of some kind.

[celest python]

there is a 3rd party open source Lua plugin that I can recommend

https://github.com/rdeioris/LuaMachine

or if you trust Epic enough, wait for Verse 😄

[zealous sierra]

That sounds like a good start! +1 for nostalgia, too 😄

[shrewd pumice]

Heyo,
What type of code should be handled by the ai_controller?

So far, I've mostly used it for interpreting perception data and altering BB keys. Am I missing something? Let me know what you guys use it for so I don't stay a confused ignorant noob.

[celest python]

Best thing about it, in your case, it has slate/widget classes that lets you create widgets runtime and players can write lua code and even get syntax highlighting

so you can just create an empty map and spawn that widget and let people run their lua code

your only problem will be designing your codebase properly to expose things to users

[celest python]

Controllers are identity of "players" or "bots", AI controller is something designed with the mindset of "bots should be treated as players" from Unreal Tournament times

You can handle any logic in AIControllers that is specifically related with AIs if it doesnt fit to other things

[shrewd pumice]

Right. I've just noticed that there seems to be overlap between what you can do in the ai controller and on the behavior tree. I know the controller runs the behavior tree, but if I can make something in the behavior tree using tasks and services should I prioritize that? Am I wrong?

[celest python]

exactly, though AIController runs BT through "behaviortreecomponent"

it just has a helper function to spawn it and init

you can slap that comp to any actor and run your BT without controller

"bots should be treated as players"
this mindset is a tech debt tbh, btw 😄

[zealous sierra]

Writing lua code with highlighting and everything sounds perfect.

[shrewd pumice]

It seems the behavior tree is the most important factor here and I should double down on it huh?

Or I should just do it dirty cause rent is coming up.

[celest python]

if your goal is to get familiar with UE AI framework yeah start with %100 getting the grasp of BTs, then you'll end up with questions of "how to implement X in best way, should I do it in bt graph or use tasks or any other solution" etc -- then you'll need to read source to develop an approach to solving common problems

also there are people who are very experienced and some of them even working on AAA games, so you wont have difficult times reaching information, especially if its UE framework related

Using subtrees help to decouple logic from a single tree

[minor pawn]

Does someone know how to set up the navigation so that you can create a rts?
My units are moving weird and pretty much shaking all the time, like they decide to move left or right every frame.
I've followed the tutorial in the documentation to avoidance to setup my units and their controller.
Also they won't stop moving, when they have reached their destination because they want to stand all on the exact point. I'm not sure if I need to create a different point for every unit, but I don't really want to create a formation. Something like in SC2 would be preferred, but maybe they also calculate the position they need to go. Also I never was able to "allow" units to push other ones away

[celest python]

how many pawns you are aiming to reach?

[minor pawn]

you mean how many at a time to control?

[celest python]

yeah, there is detour crowd manager thing in Unreal Engine, that gives you avoidance, but it doesnt scale well

like they decide to move left or right every frame.
this smells RVO avoidance, which doesnt work well as detour crowd

[minor pawn]

on my current test, i've controled like 4 to 6 units

[celest python]

if you want your pawns to keep their formation while they are moving thats a lot of work

UE doesnt provide that out of the box

[minor pawn]

I don't really need that 😄

[celest python]

then I think you can just use EQS to create formation at move goal and use detour crowd

[minor pawn]

I'm using RVO with the detour... as told in the tutorial right from unreal. They don't go well together?

[celest python]

yeah they dont

use detour only

though I also thought they go well together but another very experienced person here warned me about it a while ago

ohh btw. I'm sorry. I just realized I completely misread your question

i thought you want formations

but anyway yeah your current issue seems to be related with the params you are using with detour crowd and RVO

navmesh itself shouldnt be a problem

also detour crowd has a limitation like target movement goal should be at max. 256 poly distance is navmesh -- based on it's source code

[minor pawn]

check this out

I can't really believe, that this is the default way, unreal ships their ai navigation

[celest python]

that's not the expected behavior

so lets start with obvious ones
1- are you using detour crowd path following component in your AIController?
2- are all AIs registered to detour crowd manager?
3- have you tried tweaking params in project settings for avoidance?

[minor pawn]

Does the Nav mesh look correct? That's while playing

[celest python]

no, agents shouldnt block navmesh

[minor pawn]

ARTSPawnAIController::ARTSPawnAIController(const FObjectInitializer& ObjectInitializer): Super(ObjectInitializer.SetDefaultSubobjectClass<UCrowdFollowingComponent>(TEXT("PathFollowingComponent"))) { }

2: I'm not sure what that means. all units have the same controller class

3: yes

[celest python]

can you try printing SimulationState on your detour path following comp, to ensure its set to Enabled

[minor pawn]

how? 😄

[celest python]

get your pathfollowing component in aicontroller, cast it to UCrowdFollowingComponent, then use IsCrowdSimulationEnabled()

it will return true if its set to Enabled

[minor pawn]

on begin play?

[celest python]

yeah

gotta leave for a while, might not be able to answer soon, but let me know with the result

[minor pawn]

I'm going to bed shortly 🙂 can continue tomorrow. thanks so far

[minor pawn]

is enabled

[rich veldt]

I explained why only doing a dot isn't reliable because it filters out items that are toward the context/target, even if they would otherwise be valid in terms of level layout (dont require going past the target), and doesn't take into account items which might be out of the dot angle (even if it's like 180 degrees) but still require running past the target to get to (like a U shaped hallway).

Labyrinthine level design sort of necessitates path check i think.. so unless you're telling me checking path points on 2-3 items once per encounter is game breakingly expensive i'll stick with it

[uneven cloud]

In the example you gave, it would pick the green one using the dot product with the settings I gave you. It doesn't filter out the ones towards the target, it filters out ones on the other side of it.

[rich veldt]

ah i misread your message. dot from the target

[uneven cloud]

Do your targets not move? Because you'd have to update cover points to handle movement.

[rich veldt]

...what do you mean?

of course they move

[uneven cloud]

If your targets move, then the cover points will need to be updated if they are no longer valid. So you aren't just checking it 2-3 times per encounter.

[rich veldt]

AI spots a target, picks the furthest possible place to hide from that target doesn't doesn't necessitate running past said target, then runs and hides.
if the target catches up or otherwise takes their focus off running and hiding, the run and hide behavior is abandoned. if they successfully run and hide, they stay there until they are counter-ambushed/attacked/provoked, or their target walks up to their "surprise bitch!" hitzone, at which point they pop out from cover and swing at the target.

there's only one attempt to run and hide in an entire encounter.

think Condemned, when enemies spot you they run behind a pillar or something until combat starts

so no, cover point validity is only considered at the very start of a possible encounter in the game mechanics.

after that the ai relies on whether or not they're in danger based on their own real position (on the way to a hiding spot, or in position at one, either way), target distance, target threat level, target behavior, other threats in the area, sounds, etc.

[zealous sierra]

For an RTS, it's usually best to calculate a goal formation at the time of the movement command.

[rich veldt]

anyway here's why just using a dot to disqualify all behind the target won't work for any interesting map designs. correct me if im wrong:

[uneven cloud]

The second furtherest should win based on distance to the querier, or the path length that you are using.

[rich veldt]

i want the querier to always run to the furthest (path length) point from the target

checking in a fairly large radius

so i dont see how the closer one would win, since the scoring is based on how far from the target they are.

[forest tundra]

what should i do to make it count for ai to fly to where i am

ai no collision

[minor pawn]

I think I can't create a satisfying algorithm. What if you move 2 groups to the same spot? What if there is an obstacle moving into the goal-location?
So many things to react to..

[celest python]

if there are already units in the place you are moving, you make them move to open space. if they are both moving to an empty space you leave it up to avoidance algorithm

[minor pawn]

why isn't there a thing provided from ue?🤔

[celest python]

what do you expect exactly? there are already two avoidance systems you can use

[minor pawn]

I don't know... Either there is not enough information or they don't work for me

[celest python]

and AIModule is mostly abstract, other than "necessary" stuff to call it an "ai framework" - so there are many missing things especially if you are working on an RTS

[celest python]

detour crowd works for me, there is something weird happening on your side about yesterday's problem

[minor pawn]

Nice. AI always felt so far away for me

[celest python]

some studios rely on 3rd party frameworks

some develop everything from zero.. by wasting 6 months minimum

[minor pawn]

if you got the money, you can buy stuff

[celest python]

yeah, even for an indie you go through some initial boilerplate stuff to build up some AI if you need a little bit more than what UE provides when it comes to #gameplay-ai

so its normal to feel some absence of necessary things in AI module

[minor pawn]

I feel overwhelmed by AI

I need to create the player AI and I don't know how to build that up

[celest python]

you can try going through pinned messages, there are some useful links about both philosophical approach to AI design and UE framework

[minor pawn]

Gonna chew through that

[celest python]

high level approach to AI design is basic, having core functionalities in a shared base class and letting different systems manipulate it. so both BTs and Player systems can drive movement, make weapons shoot, trigger abilities etc.
AI part just requires some additional tools to "make decisions", and behavior tree is most common one. its useful to prioritize behaviors, meanwhile there are also other tools like finite state machines, GOAP editors, utility systems (which go well with BTs)
each tool introduces a different workflow, once you get the workflow and set-up your system modular and reusable as possible it becomes very easy to design things

[minor pawn]

I've created a GOAP from a tutorial 😄
I dont't understand the algorithm but I'm fine with that

[celest python]

never worked with GOAP, but i keep heard its mostly overkill for many situations from the people used it

definitely has its place though

[minor pawn]

We planned to use it in our story game, but we changed it to a scripted approach using the Flow plugin

[zealous sierra]

So the big reason to set the formation on the move action is exactly so that you can label individual spots as taken for future move actions with other units. This becomes more important in networked RTS games where deterministic commands become more important.

If you're making a single player RTS it doesn't matter as much, though.

Is there no open source RTS AI framework for UE? I should see if I can make a plugin from some of my work in it 🤔

[minor pawn]

There is one, but it's not maintained and it feels very overly complicated to use. I've build my own upon this and changed the combat system to use the Gameplayabilities.
Sounds like I can learn a lot from you

[zealous sierra]

Throw up a link to it?

I did some work on an RTS a while back. And by a while back I mean like 2014.

[minor pawn]

https://github.com/npruehs/ue4-rts

[zealous sierra]

Yoooo it's Nick!

Very cool!

I think highly of myself, but I probably won't have anything that beats this knowing who did it 😅

Which parts do you find complicated?

[minor pawn]

The fact, that every feature has its own component seems a little off to me. I'd rather merge some of these together. From a designer perspective, you don't want to switch to every component and make changes. Name, Icon and description can be placed perfectly into a single component.

The Fog of war is ok for a free to use plugin but I think it's hard to change

And sadly he has no time for it so it's kind of spoiling. Not working in UE5 for example. The AI BT Nodes are created in Blueprints which I moved to c++

[zealous sierra]

Looks like there's a UE5 branch.

I'm going to download this and give it a go later!

[minor pawn]

Which was updated 7 months ago 😅
So that's newer than my approach

It's a good start though

And he's from germany so it makes me a bit proud 😄

[zealous sierra]

How far along are you in your game?

[minor pawn]

not very far. I'm struggling with the unit control

[zealous sierra]

Just reading the code, the camera looks pretty rudimentary. One quick "win" might be to look at the universal camera and integrate that. It's free this month, I think.

AI is obviously the focus here, though.

[minor pawn]

I really like the approach on the Lyra Game for "Experience", where you can easily create different game modes.
But it's hard to "free" it from the rest

[minor pawn]

Whenever there's new free stuff, I'm collecting it, so I'm sure i got that

[zealous sierra]

I'll take more of a look when I get back to a computer! I was planning to make a miniature RTS myself this coming week, so I'll run into the same problem as you, I think.

[minor pawn]

The universal camera looks interesting

[clever skiff]

is there a particular reason a behavior tree doesnt run even tho run behavior tree is running in th ai controller

and when i look in the tree it not even showing it being ran

though the variable is being updated

[clever skiff]

never mind it didnt like nothing was in it even tho i had a service in it

[ocean wren]

There's also an RTS thingy by an AAA developer where they open sourced their abilities system.. can't remember the company name though

Dontnod? something like that

Daedelic

https://github.com/DaedalicEntertainment/ue4-orders-abilities

There's a video from Unrealfest Europe 2019

[minor pawn]

And Nick also contributed

[ocean wren]

Nick?

[minor pawn]

Nick Pruehs

[neat marsh]

I'm creating a simple realtime autobattler and have been using Behaviour trees but as I scale i'm starting to wonder if I should be using something else like the StateTree system.
My AI just needs to essentially attack move but scaling above 1-200 and my performance is sttuggling

[grave coral]

Does anyone know if we can till what the AI is navigating toward currently? I’m trying to disable RVO avoidance if the AI is navigating to a navlink because they keep cluttering up if multiple AIs go after 1 navlink and none of them reach it because of RVO avoidance

[grizzled bolt]

what is the recommended approach for setting up a large enemy mech AI ? O_O

[celest python]

depends™️
but seems like modular AI for pieces would be a good idea, like how Alien Isolation used two AIs for the Alien. one for body and one for head, and a shared memory (blackboard) between them

[grizzled bolt]

sorry, by setting up I mean like , should I be using the Character class and scale up the Capsule? right now as it is, the navigation is not taking the mesh size into account

[celest python]

are you not using a character currently?

[ocean wren]

Isn't a UGameInstanceSubsystem meant to auto-register? it appears that inheriting from it doesn't actually instantiate it