gameplay-ai

[slow bobcat]

As I see it, you have 3 options:

• figure out why your stuff doesn't render
• create your own smart nav link component
• add the proxy nav link as a child actor if your door BP (child actors are generally a bad idea)

[obtuse igloo]

So I just create the custom component class and add it to my door actor and i can just use the smart link then?

If all else fails i'll probably just make it so that AI either can't use doors or the door is a physical hinged thing that can just open and doesnt affect nav 😭

I really appreciate your reply though, i needed the insight

[slow bobcat]

Actually well... I told you to inherit etc but you can just add that component straight I think

[obtuse igloo]

I couldn't I had to create a component that came from NavLinkCustomComponent, otherwise there's only just "NavLinkComponent" which doesnt have the smart link stuff

On another note, even though I've made the new class that inherits from NavLinkCustomComponent, I cant set the smart link variables like if the link is enabled or what direction

NVM I found this;

    /** change state of smart link (used area class) */
    NAVIGATIONSYSTEM_API void SetEnabled(bool bNewEnabled);
    bool IsEnabled() const { return bLinkEnabled; }```

I was trying to look for a term more specific like SetSmartLinkEnabled lol

[keen crow]

has anyone had a problem with recast navmesh actor going crazy with rebuilding dirty areas in places where there is even no navmesh? it floods my vislog which causes FPS drop to 3-5 from 60-90 but also i think it's bad enough on its own that recast actor is doing all this every frame, but I have no idea what to do with it 🤔 I'm using world partition if it matters and I checked that physics is disabled on static meshes that are mentioned in vislog

[obtuse igloo]

to add on, i can't access ``` /** delegate to call when link is reached */
FOnMoveReachedLink OnMoveReachedLink;

[slow bobcat]

you have your own child. Just add a getter to expose it in your child class

that said, this is how you set it up
SmartLinkComp->SetMoveReachedLink(this, &ANavLinkProxy::NotifySmartLinkReached);

Whenever OnMoveReachedLink broadcast, NotigySmartLinkReached will execute

[obtuse igloo]

The smart link comp can't be accessed weirdly enough, i have to use UNavLinkCustomComponent::OnMoveReachedLink

ah

[slow bobcat]

that's your component

SmartLinkComp is the UNavLinkCustomComponent the ProxyNavLink uses

[obtuse igloo]

OH LMAO hsdjkfsdkfj man running my brain at 4am is just not it

[slow bobcat]

in your case, it's your custom implementation

go to sleep. Resting is your best resource when programming

[obtuse igloo]

Yeah haha Thank you lots for your help again 😵‍💫

[final prism]

Hey has anyone here encountered the warning on the State Tree Reference that outputs when it calls "conditionally sync parameters" and the parameters aren't synced at that point? I tried tracking down where it should be syncing before then, but the only occurence I see is a deprecated call in the state tree component? Am I missing something?

Well there's one more occurence when the State Tree Reference itself is loaded in the editor for editing, but that isn't guarenteed to happen 🤔

[slow bobcat]

not sure what could happen but rebuild happens for two reasons:

• a nav relevant object is added or removed to the map. This could be an actor or a nav modifier volume for example
• Something that can affect the nav moves.

If you eject whille playing and show the navigation, do you see that part going red or anything? That part on the right, is that a nav mesh bounds volume? if it is, why isn't there any nav?

[keen crow]

If you eject whille playing and show the navigation, do you see that part going red or anything?
that's the point - there's no navmesh anywhere near the region where vislog reports dirty areas happening. There's only 1 navmesh bounds volume which I showed on 3rd screenshot, all the other areas are just game-related volumes that do not affect navigation in any way and neither is there any nav modifier volume.

[final prism]

probably some shennanigans with the world partition navchunks

[keen crow]

navmesh isn't world-partitioned though 🤔 at least I did all I know possible to have it not being affected by WP

[final prism]

Out of interest, how do people deal with the implications of potential data duplication by using blackboards? Say I store something like a list of locations for a patrol route on the ai controller and want to access those for a move-to on the behaviour tree, seems like I would need to duplicate those locations onto the blackboard? Or is there a way to guarantee the 2 stay in sync?

[slow bobcat]

You can create your own bbk type as a struct and put the array of points inside

[final prism]

Would that be a reference to the original data or still a duplicate?

[slow bobcat]

The other thing is "where from the BT do you need to access it? Because if it's a decorator, task etc... You can just call a function in the AI Controller that returns said points from within the node

[slow bobcat]

All the data would be your bbk

[final prism]

Yeah mostly about reusability of existing nodes with said data, otherwise you'd need to write a duplicate node for every different way of accessing data

[slow bobcat]

Then just have a function and call it from the nodes code. Unless you need some logic that can take different arrays of locations, no need to have it in a bbk

What I mean: BTTask_FollowPoints and a designer can pass an array of patrol locations or an array of last know player locations or an array of "steps visible on the snow within X meters"

If you have that kind of logic and said point arrays can be different things up to the designer, then a bbk would make sense. Otherwise... You can just store them anywhere in code and have the task retrieve them calling the internal function

[final prism]

Also is it possible to use a blackboard key selector from an actor that isn't the blackboard? I tried doing it the same way as it works in the behaviour tree by referencing the correct asset on load but that didn't seem to quite work

[slow bobcat]

No, the bbk selectors offers you a list of bbk's owned by the BB

As very well pointed by @keen crow below, you can have BBK's in any asset implementing IBlackboardAssetProvider
A link to an example is provided below.

[final prism]

If I have a reference to the blackboard asset itself?

[slow bobcat]

Then yeah... But I'm not following. What's exactly do you want to do? Reference your BB somewhere and select a bbk from it?

[final prism]

Avoid having direct string references to blackboard values

[slow bobcat]

I do that because I have a conditions tate tree that can check values from a BB (we use both, Bt's and St's)

[slow bobcat]

You could define enums too if that's easier

[final prism]

Well the issue is if the name of a key changes you are basically screwed unless its a direct reference to the key via id

Not sure how an enum would solve that? Wouldn't it just translate to a string for the blackboard key?

[slow bobcat]

So... Yeah. This is an old problem.
The enum basically ensures 3 things:

• all bbk names are available everywhere for everyone
• you don't commit misspelling errors since you don't have to write the string.
• if a bbk is renamed, you only need to fix the enum, not every single use of that string in your project

But the inconvenience is that you need to maintain the enum

[final prism]

Ah I see, so the crux is that everyone that wants to modify the blackboard key name would need to know to also change the enum name?

[slow bobcat]

Yep. It's a minor price tbh. Bbk renaming doesn't happen that often

At least in my experience

[keen crow]

it is
#gameplay-ai message

[slow bobcat]

But you are picking the BBK from a BB.
"Also is it possible to use a blackboard key selector from an actor that isn't the blackboard?"

[keen crow]

no, I am specifying a BB on whatever I implement IBlackboardAssetProvider on to be able to make FBlackboardKeySelector work, but since it's just an interface you need to implement you can have it on whatever actor/component/data asset you want

[slow bobcat]

aaaaah
So if you implement IBlackBoardAssetProvider in any actor, you can then use it as if it was a BlackBoard (an asset with bBK's)?
ooook completely missed that

[keen crow]

you don't use it as a blackboard, you are just able to use FBlackboardKeySelectors on whatever you put it. So sometimes I have to immediately throw something from AI controller into blackboard. I have a data asset that implements that interface and contains some FBlackboardKeySelectors which I use to get BBK name instead of hardconding it or having in some named constant

[slow bobcat]

yeah, moddified what I meant.
That's cool. Funny how I never realised that

[final prism]

Ah the trick was inheriting from that interface, i guess there's some editor code that runs somewhere to populate the keys

[final prism]

Why is the blackboard key type for objects non-const? 🤔 Its not like its modifying the underlying object (hopefully)

[low sonnet]

Is it a bad idea to have thousands of smart links? (I can’t use simple links because I need my ai to jump at the start of the link)

[solid wedge]

What is the purpose of a BT Task's Blackboard Key? Just a preset variable easily accessible?

[harsh storm]

To reference a key from the BT's blackboard.

Blackboards are how you handle a decent chunk of data from outside the BT overall.

[solid wedge]

.SelectedKey

[harsh storm]

So you can more easily reuse the BT

[solid wedge]

Is this any different functionally or performant than just calling the blackboard and saying. Give me value for key "Name"

I see a lot of people storing public variables of type BlackboardKeySelector, why not just pass the "Name"

[harsh storm]

Because the BB key selector removes the possibility of typos. As well as supports refactors.

If it was just an FName - it'd be ripe for errors.

[solid wedge]

Makes sense. So if the name of a key changes, you dont need to fix all BT code

[harsh storm]

Correct.

[solid wedge]

Im seeing the vision now. Ty

So its a struct with the goodies inside. It gives you the class too. Very pack a punch.

Just need to figure out where to declare these. Would this be a top level service with public vars

[harsh storm]

Just declare them in the service/task/decorator.

[slow bobcat]

I see the problem of maintaining them, since there are lots and finding an error with one will be difficult (ie: someone moves the assets and not your nav link end is not on the ground). Imoact on performance when using dynamic nav mesh would probably something to consider if high numbers of nav links are processed per frame to build nav

[slow bobcat]

Ah teh other thing is the actor count. The more you have, the worst for level streaming. But I haven't used world partitioning so no idea how it's for it

[final prism]

short answer: yes. long answer: you will run into issues with performance due to the actor count. Unreal has recently added autogeneration of navlinks without the use of one actor per link. The way it works is by simply using a single instance that contains the logic (the cdo) and just run that logic with additional info provided by the start/end locations of the navlink and the info of the actor using it, its a bit limiting but a lot more performant

[mystic oak]

are you guys using behavior trees or no? asking because im not sure there is any gap in the same way blueprints have gaps that only c++ can fill

[winter moon]

Any ideas why my navmesh isnt generating or navigation is showing? The setting is literally blurred

[low sonnet]

I built a tool to automatically generate nav links everywhere I want, so maintenance shouldn't be too bad at least

[low sonnet]

Exactly, actor count is the thing I'm worried. I did look briefly at the automatic nav link generation in 5.5 (I can't use it since it's too simple, my own generation tool works better for my use case) but I'll look a bit more at the code to see how they handle multiple smart links per actor, thanks

[robust veldt]

DDoes any one know why nav mesh is not complete in that flat area, no collision, no nav modifier. Command RebuildNavigation has no effect.

[vagrant python]

Having an issue with using the trace test env query, it doesn't seem to be working. Trying to create a query that generates cover by finding a spot not visible by enemy locations, but a query to find a spot not visible by the querier isn't even working. Any reason why?

This query only returns the furthest point, it doesn't cull points visible by the querier

[slow bobcat]

Aaah is that how it works? I thought it was adding actors! Oooh that's neat. I got a request to check this from the design team just yesterday

[slow bobcat]

Absolutely everything in Unreal has gaps that can only covered in c++. In the end, it's a c++ engine with a visual scripting layer on top

[slow bobcat]

What's oyur setup for it? How are you generating it?

[slow bobcat]

Do you have levels below or on top?

[keen crow]

try setting bool value to true in trace test params. its bool value means "do you want the trace to hit something from item to context", so in your case you have it set to false so it will provide you points where there's nothing blocking from item to querier, and from what i understood you want the opposite

[robust veldt]

That level is a sub level of empty level.

[slow bobcat]

not sure I understand. The image you show is within a sub level with nothing but a floor in it you mean?

[robust veldt]

Yeah, It has some collision for boundary but it is correct.

[slow bobcat]

and the nav we are seeing is all within the collision boundaries?

[slow bobcat]

also this doesn't answer my question: are there other levels below or above the one we are seeing?

[robust veldt]

It is not inside. But surround play area.

[robust veldt]

I am checking.

[slow bobcat]

this might be the issue. Difficult to say without seeing more of it with the collision volume visible

[robust veldt]

I don't see anything above it.

[slow bobcat]

and below? just to confirm. Unless you have verticality in your game (levels are on top of each other) ignore my question.
One test you can do is delete objects and rebuild paths. Delete one, rebuild, see if anything changed, continue.
It's a crudre way to deal with it but...
You could also run "show collisions" command see if any collision matches the shape you are seeing

[robust veldt]

Our game is top down so, there is no level on top other.
I already check the collision. No collision show up in that area. And this issue only happen on build.

How it look on UE.

[slow bobcat]

ufff I had this problem in the last game we released and it was a pain to figure out.
We had to hack a bit the engine to print the volumes of the bounds considered for nav building in a build, dump them in a json and load them in editor through visual logger.

What kind of nav do you use? I'm guessin Nav with modifier only?

[robust veldt]

We use Nav bound volume. I also check the bound volume, and it is cover whole area.
The actor for boundary has nav modifier to avoid nav mesh is gen inside collision.

[slow bobcat]

and you don't have this problem while PIE, only in game builds?

[robust veldt]

Yeah, Nav mesh in UE like this:

[slow bobcat]

There's a chance that you have nearby sub-levels that are being loaded/unloaded due to level streaming (or WP) that are affecting this. It will depend on the order.
What I mean is this.
You can probably repro it in editor if you load/unload levels in the same order you do in the build
https://github.com/mbrpistoni/UnrealHowTo/tree/main/FixForOverlappingStreamingVolumes

the videos in that github explains what I mean

it talks abour overlapping streaming volumes, but that can happens with nav modifier volumes that are shared between two sub-levels

My guess is that some sub-level is streaming out taking the nav generated in shared tiles with it

[robust veldt]

Thank a lot. I will take a look.

[robust veldt]

I will try to turn on sub level to check if any level match that area.

[slow bobcat]

a quick test is to play in the editor the same way you would play in the build, so you can cause the sub-level streaming to happen in the same order

and print in the log which ones are streaming in/out

[grave sequoia]

The capsule is always squeezed out
Is there any way to prevent the character capsule from being squeezed away?

[slow bobcat]

What? What do you mean?

[grave sequoia]

@slow bobcatThe character capsule encounters the AI ​​capsule and spins and flies away

[slow bobcat]

Are you using RVO or Crowd manager? How are you moving your characters? Sounds like you are just forcing forces upon them

[grave sequoia]

I was around the AI ​​and let it spin around and push me away

[slow bobcat]

Ufff difficult to help without knowing how are you handling movement, rotations etc

[grave sequoia]

I solved it by setting the AI's capsule so that the character can step on it

[near flame]

hey can someone help me with state trees?i want to make the AI choose a random point in a radius and then go to it

i read the documentation and watched 2 h of tutorials but cant figure it out

[slow bobcat]

Try making move to location a child of get location

Get location and move location, being siblings, will run at the same time. Whoever finishes first will trigger the transition logic for it and/or the parent

[near flame]

also does not work

@slow bobcat do u know state trees well?
if yes ,can we make a call today?

[slim quest]

Hi All, not too sure why my Task isn't working. It should focus the attack target (the player character) but they just seem to look ahead instead. Could this be because they are registering themselves as an attack target as they are children of the player character (Needs to be this way due to another mechanic in the game)

ok- I have found out that this is the issue, but I'm not sure how to fix it. Any ideas?

[misty wharf]

So you're saying the problem is that your NPC is targeting itself, rather than the player, because it inherits from the player character class?

[slim quest]

that would appear to be the case, yes

[misty wharf]

How are you selecting the target?

As in, where does the value in the blackboard key come from?

[slow bobcat]

Can't do calls sorry, but we can have a private chat

• how are you reading the selected location from move to?
• is the selected location in the nav mesh?
• what is it failing in the logic? Difficult to say by the picture

[slim quest]

It comes directly from the AttackTarget key in the blackboard. It works perfectly fine for the boss character (who is not a child of the player character), and with the fact that the enemies are shooting each other, this is the only thing I can think it to be

[misty wharf]

Yes I mean what is the code that puts the value into Attacktarget

If it ends up targeting itself, then that would indicate the AttackTarget value is set to itself, rather than the player, so the problem should be in the code which is assigning a value to it

[slim quest]

ok, having taken a look at it and using print strings to make sure it is registering the attack target as the player, I can confirm that the attack target is working fine.

I have no idea what is causing this now

[misty wharf]

I don't remember for sure how the focus thing works, but it could be affected by the character movement component

You might need to turn on "use controller rotation" for it to work, or some other setting like orient rotation to movement might affect it

[near flame]

i wrote u a private message

[near flame]

guys ,how can i use the team agent perception in blueprints?
i read the documentation for state trees and the section below is c++,but i dont know it

[misty wharf]

The interface can only be implemented in C++

[near flame]

hmm.so i just skip the c++ part of this document?

[misty wharf]

Maybe, it depends on how the perception stuff is intended to be used

Without the interface, you can't use the builtin detection by affiliation system to determine who's on which side

But if this isn't a problem then you don't need to use the interface

[near flame]

if i dont have a numpad,how can i turn this debugging tools?

[slow bobcat]

Virtual keyboard in Windows or remapping the buttons in the engine settings. My advice: buy a USB numpad. They are cheap, easy to find second hand and it will ease your life

[slow bobcat]

If you are rich, 8bitdo retro numpad (cheff kiss)

[near flame]

https://dev.epicgames.com/community/learning/tutorials/lwnR/unreal-engine-your-first-60-minutes-with-statetree

[near flame]

those who made this tutorial ,why does danger not work properly?
AI agent does not run when it sees threat

i repeated all like in tutorial

[stuck jolt]

Any help here would be appreciate, I am trying to learn AI and setting up a state machine. I have a find farm task that finds a closest farm that works and i have a move to task. Right now the move to task never works, even if i hard code the destination in what am i missing? I will say i have an initial setup without state machine and is just blueprints with AI moveTo that does move the NPC

[vagrant python]

For "report noise event" as well as the sight sense how would I configure it such that the strength returned is based on distance to the recipient? Right now it's just a constant 1 (or loudness for report noise)

I guess I would have to determine that after the fact during processing the stimulus

[slow bobcat]

Could it be because of the part in c++ that you skipped?

[slow bobcat]

Is the state with the taks to move running? Does the taks run? If you open the state tree debugger, what do you see?

[robust veldt]

Hi @slow bobcat do you know the command RebuildNavigation work on package build. It used to work in 5.0.3 but after upgrade to 5.5 it seem does not work.

[slow bobcat]

No. This is how the nav build works:

• Dynamic: builds on demand PIE/Builds at runtime. Either around an invoker or within affected nav volumes entirely
• Dynamic with Nav Modifiers: you can only build it PIE (command or Build Paths tool) or during cooking time if specified (this I'm not very sure since we have a separated commandlet to build the nav in our pipeline). Tiles will re-build in runtime if affected by a nav modifier. No new tiles will be added
• Static: can only be built PIE or through the cooking process (same thing, not sure if you can specify in the cooking pipeline or needs to be a separated step as we do). It never changes at runtime.

Maybe if you use Dynamic or Dynamic with Modifiers it will work, I have no idea. I never called rebuild on demand within a build at runtime. Sounds like a bad idea to me

[robust veldt]

That is how our project setup. And it is bugging me.

[slow bobcat]

but why would you call re-build navigation at runtime instead of leaving it to the system to decide when to do it?

[robust veldt]

Our project use run time procedural dungeon. I am not the one dev it so I am not sure why is the reason.

[slow bobcat]

AFAIK there are two ways to handle procedural generation:

• either you construct your dungeons like Diablo3, where you have a set of "dungeon parts" you combine, like model kits (hallways, curves, rooms etc). Each of them has it's own nav baked and you streaim in the nav (this is the same as you would do using the old streaming levels system in unreal if you wanted to stream in navigation)
• Or you use Dynamic Nav and you generate navigation around invokers (this could be the player, the AI's or certain actors that are streamed in when you construct the dungeon)

You shouldn't need to call rebuildnavigation, which will force building the navigation on the entire game at runtime. You can of course (Whatever logic that command is calling you can access in c++ at any moment), but I think it would be better to use Dynamic around invokers and let the system build just the parts it needs.

[robust veldt]

Our project use pure random, there is no dungeon part.

[slow bobcat]

random is still running over a set of assets etc
You still pick a random floor etc right?
Among that logic, how are you handling the nav? You should be bringing in some nav bounds volume no? or are you creating one at runtime and adjusting it to cover whatever you generated?

[robust veldt]

Each dungeon is in sublevel and has a nav bound volume that cover whole dungeon.

[slow bobcat]

ah ok I see. So you have an empty sub-level with a defined size and a volume and you add stuff inside. Then Dynamic Nav around invokers should work just fine

but if that's too expensive on cpu for you (it was for us for example), you will need to indeed call the build nav process before presenting the level to the player. I think Nav Generation has multi-thread support (it's definitely async, but not sure if threaded right now). If that's the case, you should be able to prepare everything ahead of time with little impact to the game thread (except for the streamin part of the level)

[robust veldt]

Our game release on mobile so that could be an issue. I am finding a method to replace for RebuildNavigation command.

[slow bobcat]

I would try with Dynamic around nav invokers and use the players and the AI's for it as a performance test. Then check with insights in a Test Build. Might be fine, it will depend a lot on the size of your level, the complexity, the invokers' radius...

[near flame]

hey ,i succeded to make the AI go to random point,wait,go to arena while a timer runs,wait and repeat

but how do i make the strafe state?

i have no idea

i will have a boss on the map and it should strafe him

AI that i worked on is marked green,and the selected collision will be the arena
when AI will enter the arena,boss will spawn and the AI will need to switch the state to strafe

[dense owl]

Focus

[near flame]

boss will also have its logic ,so i should strafe him,not the arena

[dense owl]

Focus the boss key while your MoveTo is running

Oh it’s StateTree

[near flame]

yeah

[dense owl]

I haven’t dabbled, but someone should know

[near flame]

so many people dont know 😦

my teacher does not know,everyone in voice channels does not know

[final prism]

Focus is a separate system not related to behaviour trees directly. Unreal just provided a node to set a focus from a behaviour tree, but probably not from a state tree. You could relatively trivially write your own "set focus" task

If you just want it to strafe towards the thing it is moving to, enabling the "strafe" flag on the move-to task should suffice

You can call "setfocus" on the AI Controller from any blueprint or code path you want

[near flame]

wow gets simpler and simpler))

the set focus works

[slow bobcat]

my teacher does not know
My advice about this: if you teacher hasn't work and released a game in unreal in a commercial way, look for another Unreal teacher if you can

[near flame]

look for another teacher?
haha

tell this to the university,not to me)

[slow bobcat]

hahah yeah... I added the "if you can" because of it

then... no other than to be better than the teacher

[near flame]

this dude heard about state trees for the first time yesterday when i asked him for help

[final prism]

Just dont forget to clear the focus when you dont need it anymore, otherwise itll be stuck there forever

[slow bobcat]

Tbh state trees are an Unreal thing while Bt's is an industry standard. But the lack of refreshment on his part is bad

[near flame]

state trees are relatively new ,and are ready for production ,but because its new,it means less tutorials and information online

ahah,i just started describinfg by problem here and saw that one execution pin was not connected,this one

[slow bobcat]

poor excuse for a teacher (or a senior dev if you ask me), specially in an engine where you have the code

[worthy lagoon]

Anyone know how to make AI avoid each other? currently they walk into each other

[misty wharf]

The two builtins for that are crowd avoidance and RVO avoidance, you can probably find info on them by googling those terms

[near flame]

on ai controller i have this send gameplay events
one for help and anther one for combat state

i have the event watchHP on AI character to check the percentage of hp
if its less than HP percentage variable i call help event from the controller(previous screen)

here i store the player variable in a state tree evaluator to further access its location

and here is seek help task

i dont know why ,but the transition to seek help does not work

here is the transition event

[slow bobcat]

The transitio event would only work if it's set on the running state. For things that fit the description "I want to interrupt my tree to X state", you can put said transitions in the root state, since that one is always running

[wide heart]

Hey there. I've got a topic I'm having trouble finding information about. I have an Open World environment with Level Streaming set up, and my AI actors moving around and doing stuff in it. They are loaded in and out when the player gets close and far from them, that's fine.

But here's the issue: If they follow the player far enough from where they started they seem to just unload even though they are currently close to the player.

When I search for this topic I find information about persistence on loading and unloading in world streaming, which is fine, but I think its a different topic. I'm trying to figure out how to not unload these actors when they are right next to the player but far from where they started. This is why I'm bringing this up in the AI section, but if you recommend I take this somewhere else, please let me know.

[misty wharf]

I've not done a lot of level streaming stuff, but are the NPCs loaded/unloaded as part of the streaming level? If the level they are part of unloads, it seems they would then also unload, so if that's the case, you would need to somehow move them out of the level so they don't get unloaded like that

[wide heart]

You're right, they are part of a level partition cell that is getting unloaded, but they aren't in the cell at the moment, they have moved beyond it. I don't know how to update which cell they are part of right now.

[misty wharf]

Yeah not sure about that either. My first guess would be using the Rename function which despite its confusing name changes the outer of the actor - the outer by default should be the level it's in, so if you were to change the outer to the persistent level, that could affect it. But not sure since I've not really done much with streaming levels as mentioned :)

[toxic coyote]

Hello. Does anyone know why navmesh lost when changing level? I have NavMeshBoundsVolume in outliner but there is no RecastNavMesh

[worthy lagoon]

I must be doing something wrong because I've tried both of them, I also cant find any videos or anything guiding

[slow bobcat]

Edit: basically Zomg already replied, I read it after writing this

Hello!
I have done this in 2 projects I'm different ways with the old level streaming, not with World Partition. The issue seems the same though: the owner of your actor is the level where it spawns. When the level streams out, it takes everything it owns with it, your AI in this case.
You have different approaches but, ultimately, it will depend on your game implementation.
If you want the enemies to be able to chase the player forever, you have two options:

• either spawn them in the persistent level. This is done by passing the persistent world as the owner of the actor in the spawning function.
• or you can change the owner as they go. This is trickier because it entails two things: one is to track on top of which level the AI is and, when on top of a level that is not the outermost owner... Do the second thing: change the owner. IIRC that's done in the same function where you set the name of the actor. Yep... You read that right. Unless it changed in ue5... That's still the case (I can check if you want, I know we do this somewhere in the game). Just bear in mind: that ownership change used to be quite expensive in ue4, no idea now.

If, on the other hand, you don't want your AI's to chase the player forever, you need to dive into the concept of Leashing/Anchoring. Basically design a system that controls how far form a location they can go. This is a deeper topic that entails many things like: should it be a radius or a volume? How do they come back? How to avoid exploits where the player runs just enough for the AI to run backwards back home and shoot it? How do I deal with a player far enough to be outside the leasing volume but close enough to hit the AI with a sniper? The topic is a headache and it depends a lot on the game. With that I can help a bit too.

[slow bobcat]

The recast needs to be in the persistent level always or that can happen. One trick is to add a small empty nav bounds volume at 000 in the persistent level. That way the recast object will always be present

[slow bobcat]

https://youtu.be/MwQzDkrs0fI

[wide heart]

Does this sound rational:

• Place spawners in the world that when the level partition is loaded they spawn AI Actors.
• The AI Actors are not "Spacially Loaded" or managed by world partitioning so they won't disappear when the area/spawner they are from gets unloaded.
• Use something like the Level Streaming Persistency plugin to save the state of the spawners spawned actors so that for example it doesn't just spawn another one every time it gets reloaded.
• Have the AI actor itself despawn (and record to it's spawner) when it's far enough from the player.

[worthy lagoon]

I watched this but it has no guide on how it's set up

[slim quest]

Hi all, I have no idea why my AI is getting stuck on this task, any ideas?

this image as well

[near flame]

does this tree look ok?

it just does not look right

but i dont know how would it look right

[slow bobcat]

Posted it so you can look for similar videos using the terminology shown there

[slow bobcat]

Yeah, sounds reasonable. I would add a small detail which is to spawn character when the player is nearby the spawner, not when it shows up in the world

[slow bobcat]

Maybe the event is not being called? Who calls that event?

[slim quest]

Sorry, I should have said I fixed it, but thank you!

[wide heart]

But in world partitioning the cell is being loaded because the player is nearby. Or am I missing something?

[slow bobcat]

Yeah but... How big is a cell? Isn't it massive? Could it be you enter a cell in the south to never pass the middle point toward north? Why spawning anything in the norden part?

[wide heart]

It's quite configurable. Right now I have moderately granular cells with a large loading radius so that stuff is loaded in a bit at a time rather than large chunks.

[stoic pier]

I have some issues where my agents seem to be entering non-nav areas and then get stuck, trying to find a point out. I made the agent radius even larger then the actual capsule radius to allow for this but it hasn't helped

seems given the agent radius they should never be getting that close

[chilly nebula]

Could be your steering logic pushing? Did you validate if you issue a move command outside the navmesh?

[slow bobcat]

Ah I see. I haven't use WP yet. Very curious about it, specially about how would one handle a vertical game like Dark Souls for example

[wide heart]

The grid is 3d, so things will load and unload when you get far enough above or below them.

[slow bobcat]

But they unload entirely right? What I mean is: if I have 5 in column A and I'm at the bottom, the entire column A is loaded no? Or can you control the loading vertically too?

[wide heart]

Hmm, I'm not sure the use case you are proposing. I'm suggesting there is a 3d grid of cells and the actors in each cell are loaded when a level streaming source is in range. In your example there is a vertical column, is it a single actor? Or a stack of actors? It might not matter much because when they are unloaded they can be replaced with an HLOD.

[slow bobcat]

Oh sorry I meant a column as a grid column. A vertical stack of cells in the 3d grid you are describing

Was the 3d grid always supported or is it a new thing from 5.5?

Ah I see it was introduced in 5.4

[stuck jolt]

yes the state tree is working, i have another task running to find a different object in the game that works correctly. I also can move the ai in that task using the blueprint ai moveTo. Its just specifically the statemachine moveTo does not move the npc

[toxic coyote]

Thank you, but trick didn’t work

[tame ridge]

Is anyone familiar with State Trees and why the run EQS query fails to return a valid actor? The correct Classes are assigned for the actor and controller class. Even making my own State Tree task returns the actor as invalid.

[slow bobcat]

Then I need more info to help.
How's your setup for the nav and levels?

[slow bobcat]

The task using the move to from a BP, is it using the same location as the failing move to?

[chilly nebula]

ST tasks run in parallel, so since you placed them on the same state your MoveTo is trying to run before your EQS task has finished running. Its best to have your MoveTo task on a separate state that is a child of the current one, and transition over when the EQS task returns the results.
However in 5.6 we are adding new patterns to how tasks run that make async patterns like this much easier to deal with 🙂

[pastel moth]

oh hell yeah Thank you!

[near flame]

when i run the underlined decorator only the wait 0.1 and next yellow task
if i move the decorator higher in the chain same thing

but i want to abort the whole move to arena when hp is less than 0.5%,how can i do so?

if i put the decorator on the task,it executes only this task

[final prism]

What's the typical approach to having low cover areas that require a crouch or some other change in nav agent size to traverse? I've seen Unreal has the concept of "low height areas" that can be marked on the navmesh, is that whats used for this kind of functionality?

[final prism]

A sequence node will run each of its branches from left to right in sequence as long as they succeed. If the decorator fails, so does the sequence

[solid wedge]

Does anyone use Detour Crowd Controller frequently? It works out of the box, but even with some tweaks I found online, my agents are still "sticking" to each other.
Any hints as to what Detour settings may improve this?

[stoic pier]

I'm trying to set something up similar to the crop out sample project. They use floating movement component but the pawns still stay on the ground. I use the same settings but seems the 'move to' in the behavior tree moves the pawns off the ground, even if I restrict the Z in my search nav task to 0.

[slow bobcat]

If you are seeing something like this, that's what it should look like
https://youtu.be/MwQzDkrs0fI

[tame ridge]

This was exactly it, that and it appears setting the vector parameter with the EQS doesn't update in the move to task when it reads from that variable. It is giving a value from the EQS but giving a 0.0.0 value to the move to task. I just put them together instead of separate tasks and it works now.

[slow bobcat]

#gameplay-ai message

[solid wedge]

Yes, I wish it just was more smooth consistently like RVO. But I don't like the idea of introducing the "push" mechanic RVO implements. I'll keep trying to improve Detour.

[slow bobcat]

RVO alwasy sucked for me. Pushing AI's out of nav, inside meshes... Crowd is nice. You can tweak it a bit to get more distance but it will look bad in narrow corridors. You will see if you play with it.
Takes some work to get actual avoidance with anticipation that doesn't look bad.
Crowd control is more a "do not get stuck" rather than "avoid each other"

[solid wedge]

I probably need to rework capsule size and potentially introduce velocity compensation during these moments. I implemented a custom velocity compensation system into my instant pawn possession mechanic and it worked decently.

[solid spire]

I'm developing a patrol system, where you can build splines with waypoints. And I'm wondering how to decide how to handle loading/creation with world partition? Given that a patrol may potentially cross multiple cells. And Because the actor is only associated with a single wp cell, how could I set things up so that I guarantee that I'll have a patrol that may go across multiple cells?

[slow bobcat]

Not using WP, but the way we handle this is:

• designers place a special type of actor that consist in a patrol point.
• said points are added to an array within a patrol component in the spawner actor that will spawn the AI. We also draw a spline and numbers at it so designers can see quickly how the patrol looks. The patrol component also has flags to define, loop path, start reversed and some other flow control flags
• upon saving the level, those actors are processed and internally we serialoze an array of objects that have the location and couple other things (we can tú actions like "look at X direction for N seconds" in patrol points so the AI patrolling does cool stuff to decorate it
  -the patrol actors placed in the level are Editor only, so they are never cooked or saved into a build
• when a character spawns, the spawner passes to the character that data into its own patrol component (patrol component in the spawner is a data holder, patrol component in the AI is what has the logic to get the next point, to know if an action is to be done at certain point etc. Functions a bt will use)

That way you are not constrained to levels being loaded or need to increase your actors-in-level count or have more cpu work to stream this extra actors in.

[solid spire]

Ahh, that's really interesting. So the information and the patrol points are ultimately stored in the spawner, which is itself loaded when it makes sense.

Thanks, that's very sensible.

I guess extending off that though, when a player approaches this patrol/spawner, is there a reference point (the spawner actor) that you check against to know when to activate it? Or are you taking into account the bounds of the patrol to ensure you can spawn it at the extreme closest to the player the earliest it may make sense to. Not sure if my question is clear

[slow bobcat]

Yeah we check proximity to spawners to spawn AI's when needed. Unloading is a bit more complicated since we check the state of the AI, proximity to the player and other stuff.

The patrol part: no, we don't use the patrol boundaries for anything

[solid spire]

Cool, makes sense. Thanks for sharing all that, super helpful as usual

[slow bobcat]

Anytime. Sharing is caring and all that

[final prism]

Usually you would want some kind of persistent state thats not an actor but contains some extra data about how to re-spawn streamed actors

[slow bobcat]

Depends a bit on the game implementation. Ours spawn back in idle form their spawner and let the AI to deal with the world context and it's fine. But if you need to spawn them back at the last location, or with the last behavior etc then yeah. You need soemthing.

[misty wharf]

Ugh, I was under the impression that if you triggered gameplay tasks from BT tasks, they would get canceled if the BT task gets canceled but nope doesn't happen

[solid wedge]

Is it normal for the navmash to be represented as long thin shards of glass?

[fathom sun]

Hello, is anyone used to UAITask_MoveTo in C++? I'm trying to use it, but anytime I try to start a new one while another one is running, it'll pause the first one, and run the second one.

The main problem I have is that when it pauses the first one, it forces the velocity to be nullified, making the AI start accelerating during a movement from scratch.

My AI is investigating a location using the first AI MoveTo task, spots a player, starts a second AI MoveTo task pausing the first one, and setting the velocity to 0.

void UAITask_MoveTo::Pause()
{
    if (OwnerController && MoveRequestID.IsValid())
    {
        OwnerController->PauseMove(MoveRequestID);
    }

    ResetTimers();
    Super::Pause();
}

[slow bobcat]

Can you share an image of it?

[slow bobcat]

You can't run 2 move ai taks at the same time because each of them calculates a path to the goal and you can't follow 2 paths at the same time.
The Pause thing makes little sense to me. You should simply stop it and start the new one because you will not come back to the first one, you can simply run another one when needed.

The acceleration can be two things:

• you are using acceleration based path following. That's a flag in the CMC. I'm almost sure it only affects the AI for stopping (deceleration) but worth giving it a try
• the other reason might be you have a low acceleration value and it takes a while for the AI to catch up to max speed.

Even though you can fix both, you will still have that frame of 0 velocity in between move task and it will look bad.

There are 3 approaches I have tried myself that can mask the problem

• #1. you can keep the walking animation happening X frames after the movement task finishes in a Timer. If another movement task starts, cancel the timer and skip the animation playing. If no movement task after, the timer will finish and stop the anim. This is easier done when your game uses montages (you call play montage)
• #2. Movement works in 2 ways that can happen at the same tine (all happen in CalcVelocity within the CMC): there could be a movement request that comes from the path following component (as a consequence of the move AI task calculating a path and requesting the path following component to move the AI through it) or there can be an input added to the AI (the function is called Add Movement Input or something like that). Well, one thing you can do is something similar to option #1, but instead of keeping the anim alive, you keep the movement alive by adding an input every frame at your current velocity within a timer. Same concept canceling the timer upon having a new move task etc. The call velocity will start the deceleration and as soon as there isn't any move request or Inputs (acceleration within that function).

• option #3 (my favorite): you always move to an actor. Basically you assign an actor to each AI (you can have a pool for performance). The actor is invisible and all it has is a scene component. When you want to move to a location, you place the actor there (set locsiton) and move to it. When you want to chase your player, you place the actor at player location and "glue it" to the player (this can be simoly enabling the tick in your actor and make it set it's location to the player locsiton). That way, you will only use one movement task and everything will be handled organically inside, allowign5you to swap between moving to locations and actors seamlessly, without pauses or anything, no "weird" timers to keep things alive. Also, you can get advantage of the path observer and do your stuff in the path events (like enabling/disabling the tick if the actor)

[slow bobcat]

One thing that can happen is that, if you have deep complicated AI where deciding what to do next takes several frames (it will never take less than 1 because bt/st evaluation takes 1 frame minimum, but 1 frame is not an issue), you will still need to combine option #3 with #1 or #2 in cases where you approach your target and then decide things that imply another movement

[deft palm]

Hi, guys: I want to write a subclass of UStateTreeTaskBlueprintBase in C++. Is there any repo that I can learn from.
I wonder how to create variables ,as in blueprint, letting the designer to bind specific objects to them in the editor

[harsh storm]

You wouldn't do the BP Base in C++

[deft palm]

Oh I figured out :
UPROPERTY(EditAnywhere, BlueprintReadWrite, Category = "Context", meta = (AllowPrivateAccess = "true"))
TObjectPtr<AUnitAIController> AIController;

just like this in the declaration of the subclass

[harsh storm]

Look at how something like the MoveTo works.

[deft palm]

I think it could be done to inherit the UStateTreeTaskBlueprintBase in C++

[harsh storm]

I know it can be done, but you shouldn't because you lose some of the biggest benefits of being in C++

The only reason I can think of where you should is if you want to make a new BP Base that all of your BP tasks will inherit from instead.

But if you're writing tasks explicitly in C++, you don't.

Tasks in ST are structs. If you do the BP Base, you're now doing it with UObjects. Which are heavier. If there is the concept of instanced tasks like there is in BT, you also don't get that.

So you take a performance hit for no good reason.

[deft palm]

what if I want to write some generic task like : go to the nearest thing which is "some kind".
and I can configure it in the blueprint by assigning a class to the "kind ref"

[harsh storm]

That is still a struct. You would expose the "kind" property to be bound to. Then you would have some other task be responsible for actually finding it.

Or you just use the built-in Move To task.

And don't write it at all

But if you did, you still wouldn't need to inherit from the BP Base class.

Look at how the built in tasks are written.

[deft palm]

Move To Task ?

how could I search them ? Thank you very much

[harsh storm]

It depends on your IDE, but you just search for the terms. Not by manually clicking around. Use the actual search feature of your IDE.

[deft palm]

Okay when you said not using bp based class, did you use FStateTreeTaskBase as the parent class?

[harsh storm]

I don't recall the exact name. Again, just look at the built in tasks.

[deft palm]

okay thank you very much

[harsh storm]

Why is this your favorite?

[deft palm]

class UAITask_MoveTo : public UAITask
{
GENERATED_BODY()

public:
AIMODULE_API UAITask_MoveTo(const FObjectInitializer& ObjectInitializer);

the built in moveto task is itself a UObject?

[misty wharf]

AITasks (and gameplay tasks, etc.) are all uobjects

[harsh storm]

Look at how the built in state tree tasks are written. Not AI Tasks.

For the most part, right now, you don't need to worry much with AI Tasks.

[slow bobcat]

Because it's super flexible:

• moving to a location, towards an actor or to chase it? No problem, it's all the same: move your lightweight invisible actor and off you go.
• want to do spread behaviors for your group of enemies chasing the player? Easy: 3ach assigned lightweight actor it's at a player location + some offset (Sunset overdrive did a similar approach)
• want to do a movement pattern like zig zag? Ez: calculate the points and move the actor to each point before the AI reaches the point. All handled within the same MoveTo Task, no velocity changes or stops while moving like you would have when using move to location

It's flexible and opens a door to squad movement formations and movement patterns in a very easy manner

[harsh storm]

Couldn't you achieve the same effect by lerping to the new location? I'd imagine the move to location has the issue due to teleporting more or less.

Never really noticed so never looked.

So like, if you're target actor teleports to a different place, would the issue still persist (the one this solves that is)

[deft palm]

sorry I have searched but I can not find one

[slow bobcat]

MoveTo ai task doesn't support changing the goal when it's to a location. You could handle it, build the code etc but... It's easier to always move to an actor

[harsh storm]

Well...that sounds like something that should be fixed imo.

[harsh storm]

Search harder. They exist.

[deft palm]

some clue?

[harsh storm]

"Move To" and look for stuff around state tree.

They'll generally be like StateTree*Task where the * represents the name of the task.

SoStateTreeDelayTask or StateTreeMoveToTask or StateTreeDebugTextTask

[deft palm]

I find it

True it is not a uobject

struct FStateTreeMoveToTask : public FStateTreeAIActionTaskBase

[woven sage]

Is the recommendation for complex AI at this point to use the new StateTree system?

[harsh storm]

<@&213101288538374145>

[pine steeple]

ah its fine

[harsh storm]

Yesn't.

BT is still more stable and far more resources

But ST is what Epic is heavily pushing

[pine steeple]

State Tree and BT are different though

BT is much better for general AI in game

[harsh storm]

And BTs are to be deprecated in the future possibly.

[woven sage]

I've got no experience with either - I've always just written my own stuff entirely

[pine steeple]

i doubt they will deprecate BT

[harsh storm]

Mieszko said the plan is to deprecate them.

[pine steeple]

its a key AI feature, and diff to ST

[woven sage]

But for the first time will be tackling something that's going to end up complex enough that I doubt that's a good idea 😄

[pine steeple]

😦

but BT != ST

[harsh storm]

I agree

[pine steeple]

both different

[harsh storm]

And I like BT more than ST

But even if they don't deprecate them - they'll just be on life support.

[pine steeple]

i can do BT twice as fast as i can ST

and BT debugging is so much nicer

[harsh storm]

5.6 is supposed to have a lot of goodies for ST as well

[woven sage]

Right, that's a thing. So might make sense to kick the AI can down the road until 5.6

[pine steeple]

ST confuses me sometimes though

like so many things can influence and its kinda awkward

[harsh storm]

They've come in here and asked a few times what are some of the pain points and why we may be using BT instead of ST

But async stuff is supposed to be getting a lot better in 5.6 for ST

[pine steeple]

only time i would consider ST is if performance was good

[pallid mica]

ST Debugging is actually quite fine if you use that record tool.

[pine steeple]

but when i last compared BT was still faster for me than ST for my ai

[pallid mica]

Would be nice if they would integrate ST into the RewindDebugger directly if they haven't already.

[harsh storm]

ST should be faster than BT 🤔

[pine steeple]

again it depends

my simple AI test with BT was faster than ST

but this was back in 5.3?

maybe 5.2

[harsh storm]

I also don't really like the property ref stuff all that much in ST to be honest. Gimmie my blackboard 😠

[pallid mica]

I think ST from 5.0 to 5.2 is hard to compare to BT. It seems to have gotten a lot of love in terms of UI and debugging tools up until now.

[deft palm]

I think they are similar in many ways

[harsh storm]

Yeah, ST is getting a lot of attention

[pallid mica]

Sure they are similar, they are both State Machines in the end.

[pine steeple]

i do like the visual style of BT better though

[deft palm]

yeah much better to debug now

[pallid mica]

BT is a bit nicer to look at when talking about AI, yeah.

[pine steeple]

i mean i have used ST ^ proof here lol

cause we use it for mass agents

[pallid mica]

^ that quickly becomes as easily readable as a nested struct array.

[pine steeple]

but grrr it annoys me

[pallid mica]

Yeah we had to use it for Mass too. Never used it for anything else.

And for Mass it's extra annoying.

[harsh storm]

but UBTTask_WaitBlackboardTime will keep existing until last dying breath of BT in Unreal (so roughly until UE6).
#gameplay-ai message
(Where Mieszko said BT is planned to be deprecated)

Well - 5.6 is also supposed to bring the ability for STs to still be functional without having to tick.

[deft palm]

Don't you think it is more readable than BT

[pallid mica]

UE6 might become quite the shit show

[pine steeple]

no 😦

[deft palm]

haha

[woven sage]

BTs are nice layout wise because they're dense and concise

[harsh storm]

If the claims from #verse are true as well...yeah.....

[woven sage]

ST has all the condition stuff so it needs to be turned into a vertical list which is much worse for space utilization

[pallid mica]

BTs laso work differently.

BTs have flow control nodes and leaf nodes that execute or do things.
And then with Conditions and Services on top of any node that wants it.

[harsh storm]

To be entirely fair - you can still model the BT behavior in the ST to a certain degree.

[pallid mica]

STs are nested states, where each State in the Path can execute/do things. Almost like a Tunnel (vs Bubble)

That's also why sibling states can't share input/output

At least iirc.

[harsh storm]

Have the highest priority stuff at the top of the ST. Enter Conditions can serve as your decorators that gate accessing the state. Then services are just another task on the state that never completes. Only let the states transition to the next state. This pretty much models how the BT works.

With some quirks of course.

[pallid mica]

Yeah, I think the biggest quirk for me remains what can use the Output of what.

Not that this is different to BT, but it still confused me.

Iirc it was this is wrong:

- Root
  - FindWaypoint
  - MoveToWaypoint

And it should be this:

- Root
  - FindWaypoint
    - MoveToWaypoint

[harsh storm]

Eh - I think BT's data handling is far more clean personally. Data management happens through the Blackboard. Simple.

[pallid mica]

So that MoveToWaypoint can use the result of FindWaypoint.

[harsh storm]

Yeah, but then MoveToWaypoint may not even be able to run because FindWaypoint hasn't completed yet 🙃

[pallid mica]

Yeah there was a way to resolve that. SmartObject stuff had an example for that.

[harsh storm]

Siggi showed some improvement to async stuff in 5.6 a little bit ago

[pallid mica]

I think you just notify the ST or so.

It generally started to make more sense to me thinking about the nested States as Scope instead of Children.

[harsh storm]

I'd rather be able to run the tasks on a state sequentially.

For situations exactly like this.

[pallid mica]

Yeah, that's what I expected and I was confused when my "Waypoint" variable wasn't avaible on the next task xD

[harsh storm]

Yeah, they can only access stuff on the parent states and global tasks.

And modifying the global stuff is pretty painful.

[pallid mica]

In BT, I'm pretty sure you'd handle that differently. You'd have:

- Root
  - Sequence
    - FindWaypoint
    - MoveToWaypoint

Or maybe a parellel somewhere.

[harsh storm]

Yeah and FindWaypoint would write to the BB

[pallid mica]

Yus!

That's probably one of the key differences. I also think BTs can't jump to states iirc.

Which STs can do I think.

[harsh storm]

They can but you really really really really shouldn't 😅

[pallid mica]

Well let's say it's not intended at least.

[harsh storm]

BB is definitely the most missed part from BT for me.

[pallid mica]

fwiw, you can probably add a BT like feature to ST.

[harsh storm]

Well - the global params or global tasks is what is a quasi-BlackBoard

But writing to it is nowhere near as simple.

Property Reference - that's what I was thinking of. (For the BB)

[deft palm]

does ST has something like Sequence?

I can not find a reason why they should not have one

[misty wharf]

Sequence? Like in behavior trees?

[deft palm]

yeah

[misty wharf]

Not really because the execution model differs from BT's execution model. What are you trying to do?

[deft palm]

• Root
  • FindWaypoint
  • MoveToWaypoint
    then we can easily get way point from the first task\

[deft palm]

then how do they handle the sequential tasks?

[misty wharf]

Sequential tasks which use data from the previous tasks would have to be nested in order to be able to access the data so it's a bit clunky

I think there's a property reference type which might allow another way to set them up though but I've not had a chance to look at it much so can't say for sure

[deft palm]

okay

thank you very much

[woven sage]

Oh right, Siggi's on State Tree now

[harsh storm]

Yeah - he has talked a bit about what is coming in 5.6

And it is definitely a step in the right direction.

[woven sage]

We worked with him at the startup he was at previously, so it's a weird context-shift for me 😅

[misty wharf]

Yeah sounds like it will improve some cases like what piggy was just asking about

which is kind of a pain esp with async tasks currently

[pallid mica]

Isn't that Context usable?

Iirc there was some context that was passed around. At least in Mass.
We even made our own for it at some point to have more data in it.

[fathom sun]

Is there any way to cancel all the other MoveTo tasks though? I always need the most-recent one, I will never continue the one I previously had

[mild pawn]

it is still Engine.RecastNavMesh or they changed it?

seems like not effecting the world

[harsh storm]

Unfamiliar with ST with Mass to be honest. I haven't had a need for Mass yet. I'm just talking about non-nerd gameplay stuff. So you have a global param and then in the state, you'd have a StateTreePropertyReference and that would bind to the global param you made. Then you can do some gets/sets on that.

Haven't tried modifying it from outside of the state tree yet though (I found a way to do that way back in 5.0/5.1, before all this extra fancy stuff)

[slow bobcat]

If you keep the task object, yeah. You can just have a pointer to it

[fathom sun]

And what if I don't? 😛 It's called from different states which

And the task themselves are wrapped with a latent awaiter (ue5coro)

[slender cypress]

If Task_A finishes immediately and you want to use the result in Task_B,
The problem can be solved by simply placing Task_A and Task_B in the same State and giving Task_A a variable whose Category is Output.
But this can only be used in the limited case where Task_A terminates immediately.
It's much more common for this not to be the case, so I'll write it below.

If you need sequential tasks, where the next task uses the result of the previous task, you can do the following:

- Parent
  - Child_A
  - Child_B

• Let Parent own StateParameter.
• Child_A updates the StateParameter of Parent. Once the update is complete, Child_A will transition to NextState.
  • To do this, Child_A's Task must have TStateTreePropertyRef or FStateTreePropertyRef.The built-in FStateTreeRunEnvQueryInstanceData::Result is a good example.
• Child_B uses StateParameter of Parent.

For example, in the StateTree in the image, Strafe State has a StateParameter called StrafeLocation_EQSResult.
This is updated by the FStateTreeRunEnvQueryTask that EQS State has, and used in StrafeMove State.

[slow bobcat]

I have no idea how the default task works for st's. I use my own stuff for movement

[fathom sun]

I don't use State Trees for that, I was referring to other states, e.g. different pieces of code that don't know anything about each other

[slender cypress]

Sorry, I wasn't writing about your question, but about an earlier topic (#gameplay-ai message , #gameplay-ai message).

[misty wharf]

@slow bobcat random Q - any ideas for how to deal with canceling gameplay or aitasks launched inside BT tasks when the BT task gets canceled? This seems kind of a pain to track

Current idea is maybe I can have a C++ baseclass for the BT Task which would somehow be able to track when you launch the tasks from it (eg. through the task owner interface) or just shove the currently active task into a variable and check if it needs to be canceled 🤔

[slender cypress]

This is how I keep GA and BT in sync. I haven't touched BT for a while so please excuse me if I'm wrong.

[slow bobcat]

We use gas for this. The flow is bt-task launches a GA and listens to onGameplayAbilityEnded. If the tasks ends, the ability is canceled. If the ability ends, it finishes the BT task.
The trick is that all our bt_tasks inherit from our own base class called BTTask_Action and, from that, we derive everything: movement, attacks and others

[misty wharf]

Interesting, I'll check how GAS handles this because I'm noticing BT tasks can be owners for tasks but it seems to just delegate it directly to the AI Controller which might make canceling only the ones from the BT task kinda iffy - thanks

Although come to think of it, I don't think they would ever be running any gameplay tasks besides the ones from the active BT task at the same time anyway 🤔

[final prism]

The move-to task is an ai task that is spawned from a bt task. This handles it correctly, all you really need to do is override the functions the same way it does it there. Its a bit of a pain to get running but it certainly works

[misty wharf]

Yeah BTTask_MoveTo basically uses the "shove it in a variable and cancel it" approach

I just have a couple which are implemented in BP's and wanted a kind of "don't have to think about it" solution to it :)

[final prism]

You can write a generic "run gameplay task" bttask, that has worked for me

[misty wharf]

Yeah that could be an option I guess, I just have a few which are more elaborate and do more logic

There's just a lot of little edge cases that need to be handled because it goes together with smart objects which could be removed while the NPC is using it and other sorts of fun shenanigans :)

[final prism]

I mean yeah, if you make it a base class you can always add functions to it, I have something like "run eqs then run a task on the result", but you can do it however you need it

[misty wharf]

So anything where I can have something get done automatically for me (like the tasks getting canceled) is helpful because otherwise I might forget and have bugs lol

[final prism]

You can inherit from the instance data, so you can have a base class with a base gameplay task pointer stored in it and have the base class handle all of that, while a subclass could still add instance data while keeping previous functionality

[outer thorn]

Hi everyone, I am fairly new at state trees and making bosses. I am currently making a game for a game jam with a boss fight in it, and i was wondering how to do different attacks using state trees. I was thinking of a few attacks, a basic attack, a charged dash attack, a ranged attack and an AoE attack. What i would like to know is how to implement the changes between attacks, I was thinking of some sort of percentage based system (like 50% of the times it will be a basic attack and so on), but i would gladly hear any other ideas that might be easier or faster for my game

[slender cypress]

some sort of percentage based system
Perhaps Try Select Children at Random Weighted by Utility ( https://dev.epicgames.com/documentation/en-us/unreal-engine/state-tree-selectors-overview#tryselectchildrenatrandomweightedbyutility ) is what you're looking for.
One thing to note about this is that only specifying a value for Weight will not work, you also need to add Selection Utility.

[chilly nebula]

StateTree is a lot faster than BT, its the only way we could do CitySample with the amount of characters there.
Performance is the main reason to use ST over BT 🙂

I'd love to hear more about what you like more about BTs, UX/UI is something we do want to improve in the future.

[outer thorn]

I'll take a look at this, it does seems that it will work for my goal

[chilly nebula]

You can still do this in ST with a selection on Sequence, but you'd have to declare the parameters on the Sequence state to be usable by both children states.

Good discussion here btw 🙂 more of this! (still catching up)

[chilly nebula]

yes, you can use "Try Select Children in Order". See:
https://dev.epicgames.com/documentation/en-us/unreal-engine/state-tree-selectors-overview

@harsh storm Can you tell me a bit what exactly you love more about BB? I actually prefer the scoped nature of property bindings, it makes the order of operations and expectations a bit clearer to me. Using a BB is more like using a bunch of global variables and there are very little guarantees the properties there will have the data you need at the time you need it.

[pine steeple]

it might be under certain situations but i tested a very simple AI and back in 5.2 (i think) the BT was faster

but it might work better now or maybe it works well cause CitySample used mass processor and it was all batched together?

[chilly nebula]

A lot can factor into the perf yeah, including how much you use tick in ST tasks. Usually the perf limiting factor is what you do outside of the behavior, like interop with other systems.

In 5.6 you can make complete ST behaviors with no ticking

[deft palm]

Hi Siggi, meaning I have to create several states to be executed their tasks in order?

I got another question: the built in state tree tasks are not uobjects, like StatetreeDebugTextTask.
How can they be found by the editor? Since there is no reflection?

[harsh storm]

I don't mind the typesafety portion of the property bindings - but BB made it clear how to give the BT data and pass it around. So even outside systems had an easy way to give the BT data.

As for data guarantee - I'm not sold on that in regards to property bindings. Because the data you actually need may not be filled out as well. As in, like the move to vector being filled out. It isn't guaranteed that the property on the actor is filled out by the time you need it. It just exists. Same with the BB in my eyes. It just exist. Timing for both systems still come down to the dev. Unless I'm not understanding what you mean by this. (Maybe you're talking about binding from one state to another state - where as I'm talking about the global params, as that is the closest ST has to an actual BB)

[harsh storm]

They are USTRUCTs

So there is absolutely reflection

[deft palm]

Oh thank you very much

[deft palm]

Excellent explaination Thank you very much

[pallid mica]

Makes sense. It's not so much about how to achieve it, but more about how the tree works vs how the tree works (or rather how I expected it to work originally.

[chilly nebula]

They are all structs that inherit from a base FStateTreeNodeBase.

[deft palm]

Thank you very much

[chilly nebula]

correct, the data guarantees of the bindings is broken when doing async work currently

[chilly nebula]

Im with you on that, we are working on improving some UX there to make it more intuitive and in line with what people first expect

[deft palm]

but The state tree sequence by order would guarantee the order of execution of async work, I think this is good enough

[chilly nebula]

correct 🙂 but as pointed out above, you need separate states setup for it

[deft palm]

Okay

[chilly nebula]

in 5.6 you can also chain tasks together 😉 almost like a mini task graph within one state

ie triggering a task when another one is finished

[deft palm]

Even better

[misty wharf]

ooh

that sounds exciting lol

I wonder when is 5.6 coming out then hmm, I was just thinking I kinda need to redo at least some of my AI because it's getting kinda janky lol

....although it's probably gonna be broken and buggy at 5.6.0 so gonna have to wait til .1 or .2 as usual :P

[deft palm]

Have some faith😅

[misty wharf]

I'm speaking from experience lol

the .0 releases always seem to have a few weird bugs but they do tend to get fixed pretty quick

[deft palm]

Good to know

[pallid mica]

The hell did I even write there. I meant to write:

how the tree works vs how the user expects the tree to work (or rather how I expected it to work originally).

Note: Don't type stuff on your phone if you haven't had a coffee yet.

[chilly nebula]

its fine, I got your meaning 🙂

[harsh storm]

Most likely UE fest or right before it.

So about a month

[misty wharf]

Good guess :)

[harsh storm]

A lot of the stuff Siggi has been saying is coming to ST does indeed sound really nice.

[misty wharf]

@chilly nebula btw do you have any knowledge about the smart object system? I need to support multiple activities from same slots, where each activity has its own behavior (eg. "take a thing from this" vs "place a thing into this" as separate activities)... and this is kind of clunky to support right now. So I wonder is this something that nobody thought of, or is there some actual intended way for this to work because I'm not sure :)

[slow bobcat]

I know we do some stuff for this but that's the other AI programmer in the gsme. He's coming back from holidays tomorrow, I can ask if you are interested

[misty wharf]

Sure, that'd be great

I know you can have per-slot activities and behaviors which is one solution to this, but it feels not so ideal because if you have multiple slots where the object can be used from, you would essentially have <number of slots> x <number of possible activities> slots in the object at the end, and you would need a solution that claims all the slots in the same position when one of them is being claimed/occupied which seems a bit complicated as well.

So instead I have some custom things I've built around it to make it possible with one set of slots... but would be good to know if there might be something I've overlooked, especially if there's an "intended" way for it to work so any future changes to the system won't be problematic

[harsh storm]

I still haven't used the smart object system 😅

[misty wharf]

It works fairly well other than the multi-activity thing

Another small thing with it is that the builtin tasks and some functions also feel a bit like "here's a basic example" type of deal and to really fully take advantage of it, you have to make your own ones

[slow bobcat]

Sounds like anything in unreal then

[misty wharf]

Haha, maybe :)

[slow bobcat]

We use them in our patrol system to do stuff at certain patrol points. Also for what we call "first encounter" which are scripted actions for the first time a player encounters enemies in certain spots

[misty wharf]

I'm making a video store simulator so as a general system it fits it quite well. A shelf with videos is a SO, which can be interacted with in different ways. For example, customers can browse or take a video from a shelf, and employees can place videos on them. A cash register is a SO, and employees will use it to process customers in the queue, etc.

[slow bobcat]

Is this a game or a proper simulator?

[harsh storm]

Game

[misty wharf]

Oh, yeah a game, I don't know what would be a proper simulator for something like this :D

It's basically like one of those job simulator games but I hope to make it a bit more in-depth since they feel like they often fall short on that

Although there is currently a bit of a gameplay loop problem that I need to solve, or change it to a more traditional top down tycoon style game which would immediately solve the gameplay loop problem :P

[harsh storm]

Don't you hate when that happens?

"Man - if I was just in a different perspective...."

[misty wharf]

lol yep

[slow bobcat]

Oh I remember this... You posted pictures once about it

[slow bobcat]

"I can solve it with a bit of transparency around the character"

Shit storm

[chilly nebula]

I dont have a clear answer for you no, but Ive seen teams build something on top to support this. I think most just do the duplication approach you mentioned. I guess in your case you want the reservation to block other types, so you dont try to take something out of a container while someone else is putting something in.

[misty wharf]

Thanks :)

And yep that's pretty much it

[solid wedge]

Is this a normal shape for NavMesh separations?

[charred dome]

Hey everyone I've been testing some stuff with state trees and so far I'm really liking them! However I haven't been able to find equivalent functionality to "Inject Dynamic Subtree"

The closest thing that I've found is a component level ability to override linked state trees by tag. However this wants static asset references in a dictionary and looking behind the curtain it seems unsafe to alter at runtime.

Is there a way to dynamically inject state tree behavior, or is there some decidedly different pattern to utilize to dynamically inject behavior at runtime? I've previoulsly handled things like "dynamically equipable weapons" by having a weapon optionally provide a dynamic injectable behavior. For example if an enemy was unarmed and saw they could get a pistol the pistol would provide them a behavior that describes how to position and use this item. I was hoping to do something similar here.

Currently it looks like I would need to handle all cases in a monolithic state tree or I would need to swap out the entire state tree with the new weapon (including ambient and non combat behaviors).

In this case am I best served just going with behavior trees or is there some way to get state trees to behave in the way that I desire?

[ruby shard]

yea, its the poly under the pawn + avoidance radius and ...

anyone could guess why my navigation raycast (the black line) returns hit (there is not wall in front of it). retuned TimeOfImpact is 0.999999999
it happens ranmoly in weird cases. im raycasting to the center of edge/portal

[livid scroll]

Why is this not working?

[livid scroll]

My chicken doesn't move at all. Also on begin play of my chicken BP I have my logic started via the ai controller bp

[harsh storm]

1. Your tasks are backwards. You have them running away and then finding a location that they should run to. So even though this is running in parallel - it reads very odd.

2. Your Move To Location or Actor node has no goal vector or goal actor to move to.

3. You're not binding the random location from the 3rd screenshot to anything in the state tree

So overall - nothing is even hooked up really. I would also advise to split out the tasks to be two sequential sibling states. So you find the location and then move to it. Easier to follow the flow of logic and more reliable of an outcome of behavior. I am also curious why you're not using the built-in move to task.

Even though it is extremely simple - but finding a random location that the AI can actually reach. Pretty sure everyone has implemented this at some point 😅

[hallow compass]

And exporting the c++ state tree AI components

I remember having to copy paste it myself to make a c++ subclass

(Also having a sort of global theme colors would be nice, creating the same theme on all my state trees asset is kinda annoying)

[fathom sun]

I ended up unchecking RequestedMoveUseAcceleration and setting up a longer (non-zero) interpolation time on the walking blendspace, which pretty much masks the 0 velocity on task switch

[fading shale]

I have a little monster character that moves around from target point to target point...I added a timer that makes him stop every 5 seconds and face the character and shoot. I've been able to get the shoot action to happen (even the rotation, which flashes quickly) but nothing I've tried regarding the ai controller or the movement component (floating pawn movement) can seem to cancel the "move to location or actor" node...

here's one of the things I've tried.

[slow bobcat]

Yep, very well known "bug" in unreal. I had this in every project. My advice is that you do what I do: start your nav traces 1 unit away from the nav edge.

[dark agate]

Hi! I want to run an eqs query in a class that inherits from UObject (in blueprints), but the node RunEQSQuery is not available. Does anybody know why?

[misty wharf]

Nodes which require a world context won't work in UObjects unless your UObject implements GetWorld

There's an explanation plus example here https://zomgmoz.tv/unreal/UObject

[dark agate]

Thank you so much! You're a life saver, it works now!

[misty wharf]

np :)

[strange tusk]

Hello, I am working on a project which is I need navigation mesh around 50 character. Their duty is just pick random location and then go there, then repeat. When I spawn all them of them performance is very poor for my case

they are not doing anything other than walk around

is there any flag or option I can play for tweak performance (I am okay with tradeoff with nav mesh performance)

[misty wharf]

Use Unreal Insights to determine what is causing the performance problem. Chances are it isn't navigation

If you have 50 characters, possible reasons include CharacterMovementComponent and skeletal animation

[strange tusk]

When I disable navmesh collision it works well

[misty wharf]

Where is that setting? Never heard of it :)

[solid wedge]

Is navmesh generated repeatedly inside of a packaged project or is it cooked in the level

[obsidian gyro]

Hey there ! Im going crazy with something and before I yeeted my project out of the window i thought i would ask here: So my ai in the landscape with navmesh doesnt move anymore, but if i put it in a level instance move, anyone know what the problem could be ?

Mention: When placing it works, When spawning it doesn't, AI seems to have a controller in both modes, AI works on a simple test level

[compact sequoia]

Hey, I'm wondering do you guys have any method for merging/diffing state trees? Outside of diffing the raw serialized conent? (I'm currently using 5.4 and can't really make the jump to newer version).

[misty wharf]

Cooked by default but it can be set on dynamic from project settings which allows it to be generated at runtime or changed at runtime

[deft palm]

Hi, Have you guys ever send a FStateTreeEvent to a running statetree in c++?

[solid wedge]

I have a basic Behavior Tree set up. Each AI has a Behavior Tree tick happening, but what is the ActionsComponent tick doing?

[solid wedge]

AI Controller > Get Current Path Points is only returning the start and end point. Does crowd detour break this functionality?
^This appears to be the case. Visualizing nav paths + detour avoidance appears to be unsolved.

[slow bobcat]

Check with Visual logger what is going on. The navigation and path finding categories should have some clues

[slow bobcat]

I do. What's the issue?

[deft palm]

It is more about GameplayTags
https://dev.epicgames.com/documentation/en-us/unreal-engine/gameplay-tags?application_version=4.27
Did you config the gameplay tag for FstateTreeEvent in the project settings like the link above?

I think If I put all the gameplay tags in here, there would be a big fat configuration here.
and it is difficult for people to understand and to use

so How do you manage your gameplay tags?

[slow bobcat]

we can add tags from code or from the editor. We use hundreds of them. Everywhre and for everything.
We have a macro function that, when you use it, look like this
CreateNativeGameplayTag(AIStateTreeEventReaction, "AI.StateTree.Event.Reaction");
That way we can use tags in code doing things like
SendStateTreeEvent(AITagsStruct::Get().AIStateTreeEventReaction);

We are quite thorough with our tag categories and hierarchy but, if we need a function to only use specific tags, we rely on the Uproperty macro to filter them
UPROPERTY(BlueprintReadWrite, EditAnywhere, Category="Discord", meta=(Categories="AI.StateTree.Event"))
that way, only tags under AI.STateTree.Event can be selected

There's also the gameplay tag filter if you have different lists for gameplay tags
UPROPERTY(EditDefaultsOnly, Meta = (GameplayTagFilter = "YourTagFilter"))

[slow bobcat]

that's just documentation on gameplay tags work. Yeah, we do that + de c++ thing. We can say we use them in every possible way

[deft palm]

You use a AITagsStruct (A unique_ptr singleton to hold all the ai related tags)
to dynamically add a gamplay tag

and use filter when you wanna choose from a huge amount of tags

Is that correct?

when using dynamically added gameplaytags, How could the blueprint of statetree configure the transition of "On Event"

because the tags are not generated by then

[deft palm]

Thanks a lot

[slow bobcat]

dynamically added tags? Tags are static data.
Not sure what do you mean

[deft palm]

CreateNativeGameplayTag(AIStateTreeEventReaction, "AI.StateTree.Event.Reaction"); did you create a gameplay tag here?

Like this guy

You create a tag at runtime

When I use editor, I could not find it

Is that true?

[slow bobcat]

no no, that's never at runtime. I don't think it's even possible

[deft palm]

Or all this kind of tags creation are written in the Constructor?

[slow bobcat]

When I use editor, I could not find it
it's a custom function we have

[deft palm]

CreateNativeGameplayTag(AIStateTreeEventReaction, "AI.StateTree.Event.Reaction"); do you create a tag by this code?

[deft palm]

I mean you can not find the tag created in code when editing blueprints in editor

[slow bobcat]

So, the way this works.
This is the code we call

const FNativeGameplayTag TagName = { UE_PLUGIN_NAME, UE_MODULE_NAME, FName(TagString), TEXT(""), ENativeGameplayTagToken::PRIVATE_USE_MACRO_INSTEAD };```

This is how our struct look like
```struct AITagsStruct
{
  CreateNativeGameplayTag(EventTriggerAIAttack, "X.X.X.X");

  static const AITagsStruct& Get()
  {
    static const AITagsStruct tags;
    return tags;
  }
};

#define GAITags AITagsStruct::Get()

So what happens is that, as soon as you use any of the tags in code, the tags will be added if they don't exist.

[slow bobcat]

yeah, if you add the tag in Editor, you can't find it in code.
When we find that case, we add the line to create the tag in code. If it exists already, it's fine, it will not be created again.
If it doesn't exist, it will create it

[deft palm]

Elegant

I will use the code 😁

[slow bobcat]

go for it

[deft palm]

May I

Thanks a lot

[slow bobcat]

I wouldn't share it if you couldn't.
Sharing is caring.
Just ask if you plan to give a talk about it. I want to mention it in a talk in September (if selected)

We have several of these structs depending on the feature/system

[deft palm]

No I will not give a talk

I just want to use it in my code

Looking forward to your talk

[slow bobcat]

will see if it happens. Scares the shit out of me to prepare it.

[unreal briar]

Hey guys,
I created a custom AISense for combat so each time there's a battle (using detection radius) it will trigger this sense.
Everything seems to work but I don't want this sense to be received more than once for each battle.
I tried adding

NotifyType = EAISenseNotifyType::OnPerceptionChange;

Which partially working, it does call it once but that's it.
Can someone explain how does the AI sense knows the perception is changed? by what parameters?

[misty wharf]

If you haven't looked at the sight sense, perhaps that's worth looking at, since I think it's the only builtin sense where the perception state can change, everything else is just "this event occurred" type senses

[chilly nebula]

Hey sorry for the late reply here, I had to fact check a few things before responding to you. You can do most of this currently if you inherit the component and add C++ methods to modify LinkedStateTreeOverrides. Just be careful you dont change the overrides during a tick.

We are hoping to get a few changes in for 5.6 that will make this easier, and ideally a post/docs/sample to show how to do this. I agree its currently too complicated to setup correctly.

[solid wedge]

I have my detour crowd working much smoother around corners, but they always try to walk through eachother when stationary. How could I add a navmesh modifier to stationary actor locations dynamically?

[broken siren]

has anyone had any success getting root motion montages to work on actors (no CMC)

[slow bobcat]

You can. Might impact performance a bit since you will rebuild the affected tile upon activating the nav modifier, so try to avoid situations where they go in/out idle constantly

[slow bobcat]

Ummm no idea if this is posible tbh. As far as I can see the root motion is translated to Movement by the CMC. I guess you will need to implement that logic on your own?

[broken siren]

I found some code to consume root motion myself but it looked really janky - like the enemies were on a conveyor belt

[slow bobcat]

But were they moving at the expected pace and playing the animation? Or what was janky about it?

[broken siren]

sort of. it wasn't perfect and felt more like a guess based on the root motion and not actual root motion

[slow bobcat]

Yeah no idea how the engine handles it low level with the cmc. I looked at it last year to fix some stuff but I immediately erased from my memory.

[broken siren]

ya, on that note im happy not using root motion anyway since this is for a horde

[deft palm]

Hi Bruno, got a question when I dig deeper to the UE' code:

why not do it like this:
in the .h:
UE_DECLARE_GAMEPLAY_TAG_EXTERN(StateMinerIdle, "X.X.X.X");
UE_DECLARE_GAMEPLAY_TAG_EXTERN(StateMinerGathering, "X.X.X.X");
UE_DECLARE_GAMEPLAY_TAG_EXTERN(StateMinerDepositing, "X.X.X.X");

in the cpp:
UE_DEFINE_GAMEPLAY_TAG(StateMinerIdle, "X.X.X.X");
UE_DEFINE_GAMEPLAY_TAG(StateMinerGathering, "X.X.X.X");
UE_DEFINE_GAMEPLAY_TAG(StateMinerDepositing, "X.X.X.X");

[slow bobcat]

If that's what you like, go for it. There are no wrongs here.

[deft palm]

Okay Thank you very much

[slow bobcat]

how do you grab the tag later? is your macro creating the variable as a static/global variable in the same header?

[deft palm]

This is the declaration in AIEventTags.h:
#pragma once
#include "NativeGameplayTags.h"

// The miner's AI event tags
namespace AIEventTags
{
UE_DECLARE_GAMEPLAY_TAG_EXTERN(EventResourceEmpty);
};

This is the way I use it:

if (StateTreeAIComponent)
{
    //Send an event to the StateTree
    StateTreeAIComponent->SendStateTreeEvent(
        FStateTreeEvent(
            AIEventTags::EventResourceEmpty));
}

and definition is in a cpp file:

#include "AIEventTags.h"

// miner's AI event tags
namespace AIEventTags
{
UE_DEFINE_GAMEPLAY_TAG_COMMENT(
EventResourceEmpty,
"AI.Resource.Empty",
"The EventTag that notifies the resource node is empty.");
}

[deft palm]

I have to include the .h that declare the tag like "AIEventTags.h"

[slow bobcat]

Aaah OK you have it within a namespace. Nice

[misty wharf]

This macro seems to be Epic's suggested way to declare native gameplay tags btw - wasn't sure if you were aware of this as you said "your macro" :) It creates a variable in the current scope, so it becomes a global unless it's namespaced somehow.

[slow bobcat]

ah wasn't aware that's an Engine Macro. I'm not fully-in with GAS since I joined my current company and I'm a bit out of the loop.
Thanks for pointing out. I'm going to ask my peers why we don't use that

[misty wharf]

My project declares them in three different ways currently because it's been around so long lol

[slow bobcat]

yeah I think we might be suffering the same issue. The project started 4 years and something ago. The first one at least. The second is being built on top of the first so... these legacy things will happen

[slow bobcat]

I asked just in case, but it seems that we just want to declare things in the .h, so we don't have to also write the definition in the cpp with another macro

[misty wharf]

Yeah I guess that can make sense

[deft palm]

True

[warped solstice]

I need a tip how to implement sometihng like Stun/Stutter Animation (stop movement etc) on damage with state trees.

• I have a state tree with many different states
• I have a global task that detects damage

What i think would be the best is something like Global Transition that will allow my state to go into Stun State no matter which state is currently active. I imagine that adding transisition to every of my leaf right now is not a smart solution. Addition transisition to main parent (like root) doesnt work also. So what can i do? Anyone got it implemented somehow?

[solid wedge]

Put it higher up. And dont use task for checks, use decorator or service.

[warped solstice]

decorator in state tree?

im not using BT

[misty wharf]

You can put the transition into the root state to have it be active for all the sub states

This should work I think?

At least I have a number of transitions I've used like that and they did work correctly, so if it isn't working the issue might be somewhere else

[warped solstice]

i was trying with some state under root and it didint work, but if you say it works for you then i need to make more tests, thanks, thats some good tip

[misty wharf]

Yeah it should. Transitions in parent states affect child states as well

[warped solstice]

ok i think i see the potential problem but i dont understand why its happening
it will be easier to show it

When my AI is in state MoveTo (Pink) the debugger is not updating it's state in real time, and when this state ends AI is going into the state i wanted it to (Stun - light blue/green)

[misty wharf]

I don't recall seeing any cases where a state wouldn't immediately react to a transition like this so not sure what's up with that 🤔

How are you triggering the transition?

[warped solstice]

its reacting immediately if its not in this MoveTo State, so i guess there is some problem with this state

[misty wharf]

I'm not on 5.5 yet, I wonder if they made some kind of option for tasks to block transitions because that's the only thing I can think of

[warped solstice]

yea, i know its some edge case, so thank you very much for your help anyway, i will try to test some cases and find solution

[warped solstice]

jesus

i found it

i had this one shitty thing enabled

thanks man, you got me on the right track

when this is enabled, state tree debugger will not update in rela time and it doesnt react to transitions

[misty wharf]

Ooh right, I forgot that is even there :)

Lock AILogic basically just stops all AI-related stuff from doing anything for the duration, state trees, behavior trees, everything

[heady silo]

When in doubt, Unreal has some flag that is totally unnecessary just to mess with you

[slow bobcat]

Also when in doubt, Unreal has some flag that is crucial to have enabled and that you haven't heard about ever and that is hidden in some weird settings menu

[north oriole]

Does BT depend on frame rate?I think I'm noticing some services are executed less often with low frame rate even though the frequency is specified in seconds

[misty wharf]

How low are we talking about?

It seems as long as the service's tick rate is >=1 then you would have to hit single digit fps for it to be affected

[north oriole]

About 15-25

[misty wharf]

I guess at 15fps if the tickrate is 1 second it could be delayed by 66ms

But it seems it would fire after the delay though and not get skipped

[north oriole]

Because services can only run on tick right?

[misty wharf]

Yeah they would run on the gamethread

[north oriole]

I see, I'll investigate further, thx!

[deft palm]

Hi guys, Have you done something like setting property to the binding actor in state tree?

Do I have to create the state tree task to do the job

If so, it's really annoying. Because I have to create multiple setting tasks depending on how many kinds of properties I need to set

[fierce zealot]

Hey, if you work with State Tree AITask, what do you use to store inter state - state. Do you use the Blackboard of the AI, hoping it has the keys, or is ther a smarter solution? (Like use one state to find & allocate a smart object and have another state to use this SO)

[harsh storm]

It really depends on how your tree is set up

I'm smoothbrain - so I keep my state tree kind of like a behavior tree honestly.

And that's also how I do my tasks

They are small - singular focused things

[fierce zealot]

I want to trigger state trees from BTs as Sub routines, so I have a BB

[harsh storm]

So one state would be get the SO

Another would be to interact with it

So on and so forth.

@slow bobcat Does a hybrid of BT and ST. So he may be able to shed some light on how they approach it.

[fierce zealot]

yeah, that's what I plan to do, I just wonder whats a good way to hand the SO handle from one state to the other. If I write it to the BB I can just use Fnames I think, as I can't use the BBKey things I could use for BT Tasks, so I can't bake the resolved key

[slow bobcat]

We rely on the BB because we use st's for high level decisions and bt's for low level. You can have a BB without the whole st stuff if you inherit form the black board provider interface.
The in st's you can create conditions and tasks to read/write from/to BB. It's not very difficult if you look into the Balck board decorator to mimci it for example

[fierce zealot]

Will have a look, thank you 🙂

[sleek grotto]

Hey all. Working with state trees. I have a task that causes an enemy to attack. Should I place the montage, (and with that the notify begins) inside the task, or should I place the montage in the task and call a custom event (on the enemy character) to handle the hit/trace/notify logic after the montage plays ?

[tough cosmos]

Hello! When people talk about 'Finite State Machine,' are they referring to 'State Machine' when it comes to NPC AI？Is 'Finite State Machine' the same thing as 'State Machine'？
Also, is using a Finite State Machine for NPC AI considered outdated, as it is generally recommended to use Behavior Trees or State Trees now? Thank you very much! 🙂

[chilly nebula]

Not sure what you mean here? The actor the ST component is attached to is available via the context and available as a tree level parameter.

[chilly nebula]

There are many ways to do this, a lot of people trigger a Gameplay Ability from a ST task and use that to trigger the animation and the trace logic. Depending on your needs, you usually need to time the hit trace with the animation so makes sens to do that outside of the ST (via anim notify and logic in a GA).

[chilly nebula]

The "finite" in FSMs refers to the fact that there is a finite number of pre-defined states setup.
FSMs are often used to make NPC behaviors, and State Tree is a tool to write state machines in Unreal (it can even be used for other things than NPCs).
You could even say that an enum type + a class member is a simple form of an FSM, or any kind of field + if/switch statements.

State Tree is also what is called a hierarchical/nested state machine, which means it has a parent/child setup of states and the whole path of states is active at one.

So if you have for example:

Root
Idle
Combat
Attack
Chase

Or something like that, if leaf state "Attack" is active then its parents "Combat" and "Root" are also active and can be running tasks in parallel to what is setup on the Attack state.
This setup is very useful to enable code re-use, since otherwise you'd have to repeat a lot of setup that is common between states.

[deft palm]

I am writing the custom state tree tasks in c++ for our AI.

There is a state in the AI where the miner find the resource node and set it as its target.
Then go for it.
So I have to write a ST task called "Set Current Mining Target" to use.
And There is also a state that set the Deposit site for the miner to deposit.
Then I have to write a ST task called "Set Current Deposit Site".

The 2 tasks are almost same

And There are plenty of similar tasks to implement like: Setting the current attacking target for a soilder.

[tough cosmos]

Thanks! So, in this Reddit post, 'finite state machines' refers only to a specific implementation of FSMs, which involves creating a State Class for each State in C++ and defining enter/during/exit functions within each State. It doesn't mean using a State Tree or implementing it with enums, class members, and if/switch statements in C++. Am I right?
https://www.reddit.com/r/unrealengine/comments/1eskk42/should_i_use_behavior_trees_or_finite_state/

[alpine rover]

hello all, i've been trying to do some research on this but i'm not finding anything. it's a niche subject i think. basically i'm trying to get an ai actor to navigate around the player in a smooth circle or arc. any ideas on how i would achieve this?

[slow bobcat]

Best way would be to move the AI to a point near the player. That would be your radius. Then set the focus to the player and add an input to the AI to it's left or right. Since the focus will be fixated on the player, the AI will move in a circle.
If you need the AI to follow a path, you can construct one calculating the points of the circle with an eqs and using them constructing an FNavPath. The create a move request with it and make your Path Following Component use it

[alpine rover]

thanks! i'll give it a shot!

[chilly nebula]

Yes currently this is the pattern you have to follow. The issue is that output parameters from tasks are new parameters, and you cannot give it a reference to an existing parameter to write into instead.

However, you can question whether you need to "Set Current Mining Target" and instead just bind to the output parameter of the find task directly.. but I dont know the specifics of how you are using the "Mining Target".

[chilly nebula]

There is no standard way to implement a FSM, so it can mean multiple things. All it means is you have a finite list of possible states, and you somewhere keep track of what the current state us.
Everything else on top of it is extra, but often implementations do transitions, conditions etc etc.

[deft palm]

Thank you siggi. Helps a lot

[celest python]

I appreciate you Siggi for contributing in AI discussions like this regardless of the non-ST/BT/UE topics

[misty wharf]

/** Indicates how Tags from slots and parent object are combined to be evaluated by a TagQuery from a find request. */
UENUM()
enum class ESmartObjectSlotShape : uint8
{
    Circle,
    Rectangle
};

Cool comment, very helpful

I need to add support to my smart objects to have per-behavior preconditions, which is possible by assigning each behavior to its own slot... but that means I have to duplicate the same set of three slots three times for this particular one I'm working on currently which I don't think I want to do...

At this stage I guess I'll duct-tape another custom feature into a new find smart object node that can do it via some other method, would be nice to have it better integrated but what can you do (other than an engine source build with modifications but I don't know if I'm interested in dealing with this)

[torpid juniper]

I have a little component that grabs the pathing points off of a controller so we can render the nav path:

But if the map changes, the rendered path won't match the character's actual movements:

like the path from "Get Current Path Points" will update to reroute, but the actual character will continue along their original dumb path that only gets as close as possible

The movement is being driven by an AI move to.

I'm actually ok with the character not updating their movement in light of new information but I need the pathing renderer to not show a hypothetical better path that isn't being used...

[misty wharf]

Interesting that it would go out of sync like that 🤔 I think get current path points should always return the path the controller is currently using

Maybe you can just cache the current path points when the path is initially started so you can just read them from there?

[torpid juniper]

The detour avoidance makes it dynamic over the route, possibly by a lot.

interestingly this behavior only occurs when the original goal seeking becomes possible. If the desired path becomes impossible the AI move to reroutes itself.

[torpid juniper]

well, it's dumb but I guess I cant check if the path is partial initially, and if it is, set the target destination to the partial path endpoint instead

hmm but that doesn't work if the path goes from good to partial and then back to good

ah well

[lunar cedar]

question with smart objects. i know you can make slots linked to each other, but is there a way to link multiple? for example with old school cod zombies, at the barricades it has kind of a slot order. so the slots closest to the barricades are active until claimed and are targeted first, then the next row is active until claimed, then the next row, so on and so forth. im not sure how to really handle this with smart object slots. im not home atm so my terminology might be incorrect with the question

[tough cosmos]

Hello, when it comes to NPC AI, is it okay to skip Behavior Tree and jump right into State Tree, or should I learn both?

And is it realistic to migrate what I've built with Behavior Tree over to a State Tree?
I'm really keen on learning State Tree, but it seems like there aren't many good, all-around tutorials out there right now. That's got me thinking maybe I should just learn Behavior Tree and then try to apply what I pick up to State Tree later on.

Any suggestions would be appreciated! 🙂

[misty wharf]

I think you could have a link in a linked slot, but I don't know if there's much logic builtin that deals with this so you might need to build your own around this

[misty wharf]

Has anyone done a custom Smart Object World Condition Schema? I'm having some trouble with it, it doesn't seem to be saving the schema type correctly into the shared definition used by the conditions, which then fail their initialization because they just get the default schema type

As far as I can tell it seems to assume EditableConditions has values in it in before it sets the schema type, and for some reason EditableConditions is always empty even though I have defined preconditions in the smart object definition

...as far as I can tell, the only place where EditableConditions is initialized is getting called from an OnPaste handler and nowhere else

?????

[chilly nebula]

State Tree has everything you need to build behaviors and can scale better in perf than BT, and is overall more "future proof". However you are right that there is a bigger resource of tutorials, docs and general knowledge around BTs still. That is something we are working on improving.

You are right that when you learn to make AI behaviors, you should be able to do them in either BT or ST. The approach is a little different and the UX especially, but also something we want to improve on in the future.

If you are completely new to making AI behaviors, then BTs is probably easier to get started with right now.

[spring inlet]

did someone mess with entrance/exit defintions on SmartObjects? Im curious if there's a way to get them through BP

[misty wharf]

Welp this is still broken in 5.5 and possibly 5.6 also based on scanning the sources on github a bit

Really looks like an engine bug, I guess they never tried using a different schema than the default even though the support is supposedly there, unless I'm doing something wrong which I can't identify 🤔

And to make things extra fun it appears there's also some kind of corruption now in the smart object definition which is causing something to trip during startup as it's breaking on a line in world condition query... guess I gotta remake it from scratch...

Wow, it looks like 5.5 literally broke smart object preconditions altogether? Is this real

I literally can't have any preconditions on smart objects now because it will always nag "Failed to initialize query for ... due to missing schema"

Oh okay good that one just randomly went away after a random number of repeats of "resave this, delete the preconditions, resave this, readd the conditions, resave this.."

But the main issue still remains where it doesn't seem to ever correctly initialize the custom schema 🤔

[charred dome]

Does state machine provide something to handle state or task cooldowns or is that something I'm meant to implement myself? I was looking for something similar to the "cooldown" decorator provided in behaviortree

[pure pivot]

whats the preferred way to have the player pathfind around enemies with click to move (and enemies around each other/the player) in something like an ARPG? is it the crowd manager stuff?

[chilly nebula]

Nothing "out of the box" yet, but its on our list 🙂 Currently not too hard to handle on your own with a combination of a condition, parameter and a task.

[chilly nebula]

Generally that whole topic is called avoidance. UE ships with Detour and RVO implementations (and another one in Mass).

[pure pivot]

i feel like the enemy horde pathing will probably fit well into the detour/crowd manager stuff but i'm less sure about having the player cleanly pathfind around enemies, do you think detour or RVO is the way to go for that?

[chilly nebula]

In click to move games, there is not much difference between a player and an NPC. Your click issues a MoveTo order, and the main difference is how frequently you can interrupt the last order and issue a new one.

[charred dome]

is there a quick way to set a state param from inside a task? I can't seem to find a way to bind the output

[chilly nebula]

You can use a FStateTreePropertyRef property on the task (instance data)

[charred dome]

Is it possible to get at this from a blueprint base task

[chilly nebula]

Im not 100% sure but Im guessing no, since you have to work with pointers directly.

Here is a quick sample.

Define the param like this:
UPROPERTY(EditAnywhere, Category = Parameter) TStateTreePropertyRef<int32> ParameterRef;

Then in code:
auto MutableParam = InstanceData.ParameterRef.GetMutablePtr(Context); *MutableParam = 1234;

TStateTreePropertyRef is a templated version of FStateTreePropertyRef that is a little easier to work with btw.

[final prism]

How would one modify navmesh generation to markup crouch polys vs walking polys? I think this requires engine changes as I don't see anything within the navmesh generation process thats supports this out of the box?

[scenic ingot]

https://gyazo.com/8f4ad9710abe0c47fd77531d1559226b

why is my Ai twitching like this ?

[chilly nebula]

My guess would be its animation related? Lacking blend spaces and such

[scenic ingot]

hmm im only using 2d blendspace

with 1 axis

@chilly nebula it feels like my units is not affectiong the mav mesh

https://gyazo.com/82985775e81cb157acbb62e4106362f3

shouldent it show on the navigation ?

[serene fern]

Anyone know whats going on here ?

I have navigation in my level

[misty wharf]

I have no idea what the problem in that is supposed to be

[serene fern]

however in my main level where I stream this level in there is no navigation

[misty wharf]

You need to have a nav mesh bounds in the persistent level

iirc otherwise it won't work

[serene fern]

so I have a nav bounds volume in my persistent level

[misty wharf]

Well then I'm out of information I've heard from others that I can quote :D I've not used streaming levels much, but maybe someone else can help

[serene fern]

I can see that is working, its the nav in the Lunar surface Base which doesn't seem to be doing its navigation

odd so sorted it by deleted and replacing the Nav mesh thats in the persient level

[chilly nebula]

For normal avoidance they shouldnt. You can make them do that if you really need to, but its slow since the mesh will be constantly updated.

[winged grail]

Does unreal have an local llm general planning solution?
or are we still years away from that

[true goblet]

Hey all, I have a simple request for State Tree visual improvement =)
Can someone move the task list in column formation? It is really hard to read tasks when they all in one line. And also it would be more cooler if they can be collapsed or hiding if it need to be

[dense owl]

I don’t believe so. Also you’re looking for #generative-ai

[true goblet]

Okeeeeey, I started to send a lot of bugs of State Tree =)
This is first one =)

Situation (First Screen):
I have (1) State Tree that activate another (2) State Tree.
(2) State Tree has a parameter for transfer data - Selected Unit.
In (1) State Tree I connected this Parameter and data is Valid 100%.

Problem (Second Screen):
When (2) State Tree is Activated Global Task in first launch doesn't see this Parameter (pointer == nullptr).
But State Tasks can see this pointer and can work with this data.
On Next Tick Global Tasks also starts to see this Parameter

Log In Third Screen.

[slow bobcat]

That's not a bug I think. The flow in the state tree evaluator will always tick the global taks and the process the parameters. The very first tick, the global taks will run before the params update (it has a specific name I can't recall now). If you look into the State tree Context tick you will see the flow

I might be wrong since I'm not with code in front of me but maybe @chilly nebula can confirm/deny

[true goblet]

QofL Request =))
If I have a lot same tasks - I cannot see edited name of the task in list of task names =)

[slow bobcat]

That said: yeah I agree with you that, ideally, params would be set before anything else

[true goblet]

The very first tick, the global taks will run before the params update
It is not what I expect =) In this logic the Actor pointer must be updated also after Global Tasks =(

I mean Context*

[slow bobcat]

I might be wrong but I r3call event consumption and global taks being the very first 2 ticking

[true goblet]

Yep, But Global Parameters - it's like Context - something solid and must be updated before any tasks will be run at all =)

I mean It is what I expected =(

[slow bobcat]

Agree

[misty wharf]

Ohhh finally I got that darn precondition for smart objects with a custom schema working

So it turns out... If you have a custom schema, it does not work at all if you configure Preconditions for the entire smart object

But... it works fine if you configure selection preconditions on a per-slot basis

I wonder where do I file a bug report for this...

[flint trail]

so, in 2025 and UE5.5 (soon 5.6), what's hot for NPCs AI (combat and friendly) - BTs or State Trees or something else (I haven't touched AI since 4.27 😓 ) ?

[misty wharf]

State Trees are the cool new thing but they may still be a lil rough around the edges so you're going to need to be able to figure stuff out yourself

[flint trail]

so, should I just stick with BTs ?

[slow bobcat]

We should start pinning answers/questions like this.

• St or bt? St are the new toy, Bt's are industry standard across engines
• my movement/path fails - > use visual logger to get more info

And the likes

[misty wharf]

That's partially why I've been writing stuff on my site lol

Easy to link those kinds of things from there

[true goblet]

I tried to work with BT several times and always dropped it in half way. But State Tree is the best solution for me =)

[flint trail]

I don't see whole a lot of training material on State Trees

[true goblet]

Oh. I forgot add the fact that connections with that parameters (in 2 State Tree) often crash project in 5.5

[unkempt thunder]

I don't think there's any need to throw out BT's for state trees if you already have built up knowledge and infrastructure for your BT's (ie decorators, tasks and services)

unless maybe you're also going to do Mass

if you're starting from the beginning, especially "how do I do AI" beginning then I'd say go with StateTree

though, as noted, there's a lot more out there to help you learn with BT's

[chilly nebula]

We have some, and more docs coming.

Here's a couple to get you started:
https://dev.epicgames.com/community/learning/tutorials/lwnR/unreal-engine-your-first-60-minutes-with-statetree

https://www.youtube.com/watch?v=YEmq4kcblj4

[chilly nebula]

Show/hide toggles are in 5.6, as well as search!

[chilly nebula]

I believe we fixed this in 5.6 🙂 + lot of other fixes around linked trees and more deterministic processing order for events and transitions.

[chilly nebula]

This works fine in both 5.5 and 5.6 for me, what version are you on? Did you actually rename your tasks in the details view?

[chilly nebula]

https://www.unrealengine.com/en-US/support/report-a-bug

Or Epic Pro Support (UDN) if you have it.

Best is if you can do a PR for a fix that we can review and clean up 😉

[misty wharf]

I'll consider it if I get around to having a source build at some point :D

[true goblet]

Not really. I updated the task name by function GetDescription where I added a tag name to the task name. And I suggest to use this names for task list while trying to link parameters =)

[true goblet]

That's really cool, thanks!)

[true goblet]

Okay then. Need to wait 5.6 to check all my problems =))

[misty wharf]

Hmm, seems there's been a behavior change with Group states in 5.5 state trees as opposed to 5.4

I had a group with selection behavior "none" as a container for states which are only entered via events... But this stopped working in 5.5 altogether, the states inside it can't be entered at all now

I had to switch the state type to "state" and then back to "group" for it to work, but it also won't let me change the selection behavior for it

Oh well, I guess it doesn't matter that much

This does make me wonder... what is the purpose of "None" as a selection behavior if this completely prevents selecting the state now, even from an event based transition to a child state 🤔

[chilly nebula]

Im curious about this, interesting pattern! What is firing those events, and what state did you expect the ST to stay in if it couldnt enter a state? My guess is the ST couldnt enter any state, which would go back to root and potentially even stop if there were no valid states to enter.

[misty wharf]

I didn't think of it that far, my main reasoning was "I never want these states to be selectable, except directly via a transition from another state", so I put them into a Group where the group had its selection behavior as None just to make sure the only way to enter the sub-states was a transition which explicitly did that.

If a None state can't be directly entered, I'd say my expectation would be an error when I try to save the state tree if that was going to happen, or it would be so that I can't even choose an option that would make it do that

It seems like covering all the possible ways a state could be entered might be kinda tricky so I suppose if it shouts loudly enough about it happening at runtime that could also work

As mentioned, not really sure what the intended usage for None is since it seems as of 5.5 it can't be entered and its children also cannot be entered :)

I think originally I was mostly just interested in testing the Group state type and seeing if organizing the states in this way would be helpful

[chilly nebula]

These are good points. We made many fixes in 5.6 regarding transitions and a more deterministic processing order, I'll take a look what the behavior is now wrt None

[lucid field]

Hello, I'm currently working on an enemy that's supposed to run towards you while attacking and I'm using this parallel node for that and it correctly works in parallel (enemy moves while attacking) but as soon as the move to node reaches its target (the acceptable radius I set), the attack sequence stops running and it's just the move to node running continuously. How can I make it so even when the target distance is reached the sequence continues to run? (so that the enemy would just stand and attack)

Looks like it's not only with the move to node but with any main task of a parallel

[north oriole]

As a follow up question to this: what about tasks?how often do they run and can frame rate affect it?For example this is a simple task that does nothing and finishes execution instantly. But while I would expect it to run every tick of the tree, it doesn't. The service does, but the task seems to run every 3 ticks or so, so it seems it is affected by frame rate. Is that so?

[celest python]

huh.. no TOptional support in STTs 😄

quick question meanwhile: is it possible to create "template" ST assets?

i.e., schema comes with built in global tasks or a tree hierarchy?

I'm working on a combo attack system and global tasks listen for input and send ST events to listening conditions/states to transition to other states

but I'll have to add all of them one by one to each asset

if not, I'll guess I create a template and manually duplicate 😄

[slow bobcat]

The way a parallel node works is that the secondary task runs only while the main does. Your MoveTo is finishing, ending the attack. Then, because you are in a selector, it finishes (selectors run until one child is a success). The tree exist and re-evaluates, enters your selector, move to finishes immediately because the AI is in place. That happens in a loop. It will probably start working again if, while this happens, you move the goal location/actor of your move.
The way to handle this is having a decorator checking the distance to the target. If too far that it need to move while attacking, run the parallel branch. If already at the desired distance, the decorator will fail and the selector will run the next branch, that will be a simple attack sequence (exactly the same you have in the parallel node). Your decorator should abort in the only attack branch. If the goal moves away, abort the branch so the tree re-evaluates and runs the parralel branch

[slow bobcat]

Bt's evaluate every frame, otherwise your service would never run every frame. There isn't a separated logic for services outside the tree logic. When a tree evaluates
Its logic, it goes into all the branches and composite nodes following the execution flow until a leaf can be executed. Then it stays there until the leaf finishes (or a decorator aborts).
If you feel things are running every X frames only it can happen your tree takes time to evaluate due to callbacks, decorator validations that take time to pass etc. For example: you might have a decorator checking for certain thing to happen in your game and that thing might be taking 3 frames to happen. Your decorator will execute every frame but it will not pass until the 3rd, when your gameplay thing happens.
It's imposible for us to know unless we scan your entire logic and that's something you should do

[nova osprey]

Hi ,I am trying to implement a turn based system. The idea is the ai responds to CanAct member of the blackboard being set to true, performing an action, then ending their turn. However in game the AI is completely ignoring CanAct and just spams its action over and over. .. struggling to understand why the behavior tree is running, even when I set the CanAct key in the blackboard to false before calling Run Behavior Tree in my blueprints. Been stuck on this for days any help is godsend 🙏

[slow bobcat]

See how your decorator is greyed out (blued out really)? That means things. If you hover the mouse you will see a message saying that a decorator in a root node will not run (the root node of a tree is not the one that says root, it's the one under it. Root node is just a visual cue so the user knows where to start adding nodes).

Add a sequence node below your Root node and call it "SQ_ActualRoot" or something like that. Then attach to it your current tree logic. You will see your decorator changes color to it's natural blue and things will work

The exception to this is if your tree was run as a subtree. In that case you can add decorators to it's root, since the root of the tree will actually be just ankther node in a parent tree (it's injected in the parent)

[nova osprey]

yep haha that works thank you so muchh

[nova osprey]

i read that and didn't understand and ignored it.. well im glad that only took a few days for adding one node to my tree.

[slow bobcat]

Next time will take 1s to spot. That's the important part

[chilly nebula]

Yes it's possible, and being done by large games. You have to rely on linked subtrees and control the injections (often a few levels deep) at runtime.
Basically you setup base "templates" that have defined injection points (called overrides), than you inject other templates or shared trees at runtime based on the needs of the NPC you are doing. Its common to share things like reactions, pathfollowing etc between different behaviors, and you can even run some of it in parallel (like hit reaction states).

[lucid field]

Got it, thanks, I'll keep that in mind

[deft kettle]

Hi so I am trying to create an AI that would be basically an Enemy archer that would learn how to shoot enemies that are moving continuously in a room
I am using Learning Agents on 5.3.2, also using C++.

My issue is that it will generate the first iteration results, it will run a little bit more, and will give me this error:
Agent with id x has completed episode and will be reset but has not generated any experience. I am having like 7 agents.

This is how my SetupCompletion looks like:

{
    TimeCompletion = UTimeElapsedCompletion::AddTimeElapsedCompletion(
        this,
        FName("TimeCompletion"),
        30.0f, 
        ELearningAgentsCompletion::Truncation 
    );
}

And this is my SetCompletions implementation:

void ULearningAgentsTrainerR::SetCompletions_Implementation(const TArray<int32>& AgentIds)
{
    for (int32 AgentId : AgentIds)
    {
        float& Timer = EpisodeTimers.FindOrAdd(AgentId);
        Timer += GetWorld()->DeltaTimeSeconds;
        TimeCompletion->SetTimeElapsedCompletion(AgentId, Timer);
    }
}

And basically the error will get spammed, and my ResetAgents method will get called everytime, which basically will teleport my enemies continuously.

I would really appreciate your help, really don't know how to implement this SetupCompletions and SetCompletions, if you have any advice, it would be verry apreaciated.

[cinder lantern]

Is there a way to visually debug AI State Tree ?

[misty wharf]

There is a debugger in the state tree editor and the visual logger also contains log info for state tree executions

[solid wedge]

Ok, blackboard keys feel redundant. Why have ai class variables at all if theyre all duplicated as keys.

Can I just store everything in the bb

[harsh storm]

You don't have to have them as class variables.

[solid wedge]

I figured class variables would be better to access/update frequently, but managing both is annoying

[misty wharf]

I have some on BB some on class with specific decorators that can choose based on them

What annoys me about the BB is that it becomes messy easily and it's kinda hard to tell what's what

[solid wedge]

The c++ syntax for getting/setting is not asthetic. And im lowkey concerned its less peeformant than referencing self-class variables.

[misty wharf]

Most likely the performance difference is fairly irrelevant unless you're doing it a lot in a hot path or have a gazillion AI's

and C++ is not aesthetic to begin with :D

I've not done BB lookups much there but if it's sufficiently clunky I suppose you could always make your own function or macro for it

[hallow compass]

ST power

BT sucks

[slow bobcat]

It's an abstraction. Ie: you have a bbk target. Target for your melee warrior ai is the player character. Target for your saboteur AI is a building. The BT Attack that does: got to target + attack is now reusable for both, independently of what a target is

Let alone that your different AI's might have different ways to get a target: some use perception, others use some custom logic in a component only them have...

[slow bobcat]

Depends a lot. Get/set things from/to a BB form within a BT node it's cheaper than grabbing actors/components. But the difference is negligible in most cases as very well appointed by zomg

[north oriole]

That task literally prints a log on screen. Shouldn't be slower than the service (which also prints on screen) but it runs once every 3 service ticks

[slow bobcat]

Umm strange. I have no idea what could be going on from the top of my head.
Are you printing the frame number in each print? Just to see if anything is printing twice the same frame?

I'm very curious abut what's going on.

Does this happen with a single AI and nothing else happening?

And just to confirm: the task has a Print and a finish task call and that's it right? Nothing else?

[north oriole]

A bit more convoluted. I was printing the world delta seconds in both to tell how much time passed in between calls. I also capped fps at 5 so I knew each frame ran every 0,2s. I noticed the service calls were 0,2s from each other as I expected (since it's tick is lower than 0,2s). While the task logs where almost exactly 0,6s from each other

[north oriole]

A single AI of that kind yeah, but other things were running in the map so maybe they dirtied the experiment

[slow bobcat]

I would try with just one AI in an empty level and print the frame number without capping the framerate, but your test sounds ok

[rose pelican]

Anyone know how to do enemy Ai capable of crawling/walking on walls?

[mighty hornet]

How are other people finding state trees in terms of performance? I have a relatively simple tree that works great for 1-2 agents (AI vehicle drivers) but performance takes a nosedive when I go much higher. With 7 AI opponents my frame time increases nearly 20ms confirmed by profiling. Any tips to improve performance?

[harsh storm]

Can you show the profile? Because that is quite abnormal

[slow bobcat]

For Crawling you can use nav modifier volumes on your nav to mark crawling areas and, while moving following a path, detect if the AI is on one. Then do whatever you need (set movement mode to crawling, scale the capsule, okqy de right anim etc). For walls... That's a completely different topic. Do you need path using walls?

[rose pelican]

The walls are spawned as a part of a different BP as dynamic terrain. The entire room is one object

[rose pelican]

Do the walls need to different objects or can the wall meshs be targeted separately from within the BPs?

[slow bobcat]

if you need navigation on the walls, it's complicated.
Unreal doesn't support generating nav mesh on vertical planes, only horizontal.

-> If you need to climb the walls to reach the rooftop of a building, you can have a smart nav link on the walls to connect the bottom floor with whatever is on top of the wall (assuming the top has nav, like the roof of a building) , then play the climbing animation with root motion, then use motion warping to adjust the movement to the wall length (or define very specific wall heights like small/medium/large and have different animations tailored for them)

-> Now, if what you need is to use the walls to navigate within the room, that's a bit more complicated.
One thing you can try is to have entry and exit points defined on the walls. There would be 4 points like in the image.

• find the closes entry point to a wall from the floor where your AI stands
• path to it
• from there, using the mesh topology of the wall as it was the nav mesh, create a path to your exit point of that wall.
• Grab the connecting wall topology and path from the entry point to the last.

This is difficult to do because:

• You need to calculate the exit points, which means you somehow know the direction your AI needs to move and which walls to follow. You can run a very simple A* algoritm to know it where only said points and your goal location are considered. With that, you know which walls you need to take the mesh topology from.
• Although all the code to path any mesh is in the engine, you need to convert your topology data from the walls (their mesh) to something the nav system understands. Or copy the A* algorithm used in the engine or create your own (I recommend the A* JPS which I believe it's what Unreal uses already https://gdcvault.com/play/1022094/JPS-Over-100x-Faster-than)
• you will need to string-pull the path obtained to cover cases like having a window on the walle

or you can use the Kythera plugin, that supports Surface navigation and already gives you wall/ceiling navigation
https://www.youtube.com/watch?v=g9_9E2eDgRM
https://content.kythera.ai/home

[rose pelican]

I see, thank you for your input and assistance. I will see if this can help my situation in any way. Thank you

[slow bobcat]

If not, ask again with a more detailed example. Happy to help

[chilly nebula]

hey, I wanted to follow up on this. Check this CL for an easier way to set overrides at runtime, it will be in 5 .6
https://github.com/EpicGames/UnrealEngine/commit/d5b44e9281bdcebbbb9fc79f58259a7b1af633ab

[chilly nebula]

You are probably doing something expensive in a task or interoping with components or features that cost a lot of perf. Ticks are also a bit heavy, but 7 agents is nothing.. we have an internal project with hundreds of character actors and city sample has thousands (but those are in Mass).

[chilly nebula]

Or Mercuna 🙂
https://mercuna.com/

[slow bobcat]

first time I look into mercuna prices. Bit insane for indies

No idea how that compares to Kythera, but 2.5k the cheapest is a bit much for most solo-devs

[misty wharf]

gotta be god tier at that price

although to be fair, I'd say rolling your own navigation solution is likely to take more than a few months so this price is quite quickly worth it for business customers

[slow bobcat]

Yeah. As always: how much do you know and how much time do you have?

[barren flame]

Why my nav agent only moves if "use acceleration for paths" is on? I tried everything, from recreating the nav bound and deleting the recast actor.
It would be ok but "use acceleration for paths" has its issues.

[slow bobcat]

How are you moving your actor?

[barren flame]

Move To in the Behaviour Tree. It was working before by the way, but then suddenly it stopped. Only that setting make it semi work.

[celest python]

personal opinion: kythera isnt worth it and their pricing isnt any lower than mercuna

and least mercuna has a pricing they give to you if you're an indie

kythera wouldnt answer your support requests and if you dont make it obvious that you are rich they dont contact you for sales either

[slow bobcat]

What did you change in between working and not working? What does visual logger say in the case where it doesn't work?

[barren flame]

I modified the BT, then I made it back as it was. Everything is the same. No reason to not work.

[slow bobcat]

I evaluated it for our game but didn't offer anything else we needed. But if you need 3d navigation or moving navigation (their boat example is nice) I guess it's worth checking out

[barren flame]

What's the visual logger?

[slow bobcat]

Your best friend. A debug tool provided by unreal. Check google for more info

[celest python]

last time I used they had a custom lua thingy and some DSL to handle data, and web API wasnt nice to use

they were reflecting your web elements to editor through asset creation but it was crashing the editor time to time

which is ok, bugs can happen but it wasnt nice they didnt contact us at all even to buy it

i like tech part though ngl, ORCA2 was nice to see in action

[slow bobcat]

What does Kythera use Orca2 for?

[celest python]

to replace RVO and detour i think

works well/better for zombie games

ooh this conversation reminded me of DetourCrowd.cpp

Siggi, if you can see this (i dont want to ping you), not sure if I'm wrong or not but it looks like everything could be paralellized easily(?)

I think each function that starts with step in detour can be ParallelFor or all steps can be combined within a single for loop and a single ParallelFor can be used

iirc thats what ORCA2 does

[slow bobcat]

Aaaaah that Orca! Hahah OK I thought you meant the MS AI/MLL stuff

[chilly nebula]

Often its not that easy, and ParallelFor adds an overhead 🙂 If you want something that performs better, stay tuned.. we have something new on avoidance perhaps coming in the future.. you can take a look at the avoidance in City Sample for a taste but that is Mass based only.

[celest python]

and ParallelFor adds an overhead
sure, but just at first glance code looked like its eligible for parallelization out of the box 🤔 but of course I'm more alien to the API so I'm not sure

[simple moon]

anyone have any ides how to add navmesh into a procedural generated level?

[slow bobcat]

Dynamic nav around invokers

[simple moon]

thanks ill look into that now 😄

[deft palm]

Hi guys, I want to get and set the global parameters in a StateTree task using c++
Do you know the corresponding API for that?

Thank you very much

[chilly nebula]

You bind them to inputs on your tasks. If you need to change values inside of the task, then use property refs. See example I posted here from a couple days ago.

[deft palm]

Oky

Thank you Siggi

Define the param like this:
UPROPERTY(EditAnywhere, Category = Parameter)
TStateTreePropertyRef<int32> ParameterRef;

Then in code:
auto MutableParam = InstanceData.ParameterRef.GetMutablePtr(Context);
*MutableParam = 1234;

this one?

[chilly nebula]

yes, that is an example of an int parameter and how to change the value of it inside a task

[simple moon]

is there a way to make ai move whitout a "navmeshboundsvolume"? 😛

[slow bobcat]

I think floating pawn? You can also add input to the character using the character movement component. And you can probably do a move request like the path following component dies within Follow Path

[harsh storm]

As an FYI - you don't need the CMC to do an Add Input call. That method itself is on the Pawn and it just ferries that info to the movement component.

So it isn't Character/CMC specific

But if you're not using a navmesh - then that is the only solution as far as I know. Because all of the various move tos do require a nav mesh

[harsh storm]

A new avoidance thing? And it isn't tied to Mass? 🤩

I am lovin' this attention the AI side of UE is actually getting.

[slow bobcat]

Don't tease us man.... That's cruel XD

[slow bobcat]

This is a blessing indeed

[charred dome]

This is fantastic! Are any of the other planned improvements (like cooldowns) happening to land in 5.6 as well?

[chilly nebula]

5.6 is filled with goodies 🙂 but direct support for cooldowns no, we need a bit more time to cook what we're thinking there.

[charred dome]

A current way to do this (implement a cooldown) is a "3 piece" solution yeah?

-in64 Timestamp - Store unix timestamp of last entry as a task level param
-FStateTreeConditionBase -- Check Timestamp vs Now
-FStateTreeTaskBase -- Set a timestamp as a task level param

[chilly nebula]

Yep that would do it 🙂 just keep the parameter in a place where it wont leave scope and be cleared.. might have to be a global parameter for that reason.
You probably want to add a cooldown duration somewhere, probably on the condition itself.

[charred dome]

If I stuff the value in a "Task Parameter" here it shouldn't get GC'd should it? This is an InstancedPropertyBag?

[chilly nebula]

as soon as the state that contains the task is out of scope (ie the state no longer active) it will get cleared

[charred dome]

oh wow I did not know that okay

I'll just establish the timestamp in a data store I keep myself outside this tree

[chilly nebula]

This is one of the key differences between BB and the property binding solution in ST, each state has a scope.. almost like its own mini BB.. Only items in that state or child states can read from it, and if no longer in scope its cleared from memory

[chilly nebula]

you can, but probably faster if you just declare it as a global tree level parameter.. or even on the root state

[charred dome]

If I want to make a bunch of cooldowns I use a map keyed on name? This is starting to feel messytown

[chilly nebula]

yes, that is indeed where it gets messy 🙂 if you need 10 different cooldowns, you need 10 parameters. This is why we want something better 🙂

We have an idea, but I cant get into details since things often change when implemented 🙂

[charred dome]

I'm starting to wonder if having a state tree manage higher level "agent state" and then using a BT to implement combat behavior might not be the correct soltuon for what I"m doing now

[chilly nebula]

thats a viable solution that many use 🙂 even internally many teams do a hybrid solution. But we also have large AAA teams that are doing very complex behaviors with lots of nested subtrees, without any BTs.
Basically go with what works for you now and you are comfortable with.

We really want to get to a future where ST is the tool everyone wants to use 🙂 but we have some work to do to get there still

[charred dome]

Yeah I already see the upside I just don't feel like solving some of these foundational issues myself when I know better solutions will probably be incoming soon. Thanks so much for the help you're an amazing resource here!

[chilly nebula]

and you all are a resource for me 🙂 helps us prioritize and see how you all are using the features we work on.

[charred dome]

hmmmm... Is there a good unique identifier on a state I could get from it's instance data? I can probably hack together something in my data storage solution to hold onto the timestamps and then I can keep all my logic in here and be part of the future

[chilly nebula]

not atm.. but something is coming with 5.6 😉 Do you mean the unique path? ie "Root -> Idle -> Combat" or a unique ID for the current instance running of a state?

If its the former you need, its called a StatePath and is a string. Im 90% sure thats new in 5.6, cant remember if we shipped it or not with 5.5.

[charred dome]

The latter. Looks like I check back on with 5.6 and see all the new stuff. Thanks!

[chilly nebula]

ok just keep in mind that ID will change every time you re-enter that state since it will be a new instance. If you want anything to persist between then the state path is probably a better choice.

[harsh storm]

Once I'm convinced - then you know you've succeeded in this goal. I'm stubborn. I'm old. And I like the straightforwardness of the BT.

[charred dome]

I don't think BT is intuitive for most folks it's just a tool that a lot of us are trained on. The BT pattern has a lot of implicit function that I think this avoids

[slow bobcat]

The possibility of having a g.tag as an ID for a state would be incredibly useful. And if we could do StateTreeComponent->getRunningStateID's returning a g.tag container, that alone will unlock a universe of possibilities

[harsh storm]

I think it is more intuitive than STs due to how STs work. Many devs have tripped up on the asynchronous nature of task processing and how state transitions work. Let alone how to pass data around.

I'm not talking about deciding how a task should be authored mind you. With BT, it is pretty much - all data flows to the BB and then the BT executes left to right. With decorators to "gate" branches.

With ST - there is the async task processing, keeping track of how the various tasks potentially complete (due to how the state transitions work), deciding where the variables should live (should it be global or should it be on the root or should it be on a parent state, etc...), figuring out the process of getting data from one state to another, etc...

Many more things to wrap your head around because it isn't just a basic hiearchical state machine.

Yeah - the BB being pretty much a global variable bag has its issues, but it is pretty straightforward

[chilly nebula]

In the near future we want pretty much all tasks to complete 😉 unless they are marked to never complete, and with 5.6 you now have controls over what tasks can trigger the transition and if you want ANY (as before) or ALL that contribute to trigger it or not. There is some more work to do here, but 5.6 is already so much better in this regard and just makes more sense.

I 100% agree that issues like this have made STs non-intuitive to pick up, but that is why I want to discover and fix all of this.

Regarding the scoped variables instead of one fat BB, I feel this is also one of the things that will allow ST to scale better than BBs since they are more lightweight. But I do agree that BBs are overall easier to use, and I want to find ways to improve how the data flow happens in STs.

Regarding the above - in 5.6 you as a ST task author can choose if your task will ever or never complete, and if by default it should contribute to triggering the transition or not on completion.
Then as the author you can also toggle this on/off per task on the states.

[harsh storm]

I'm really excited to check out ST in 5.6, but that is still a month away. I don't have the infrastructure to have my team on a source build.

[north oriole]

Ok it's the same

[deft palm]

Hi Sigi,How to set global variables outside the state tree.Like Blackboard’s parameters’ settings

Thank you very much

[livid scroll]

so based on this, why does my first branch of my behavior tree go back and forth between flee-patrol really really fast & my chicken never make any moves?

[deft palm]

how about only connect ai Moveto OnSuccess to finish execute node

[mighty hornet]

Of course, you were right. I switched my evaluators to run every 0.1s instead of on tick and managed to save a bunch, but all the calls to the function in green are still costing me a lot...

[slow bobcat]

This is very strange. Just to be sure:

• you have one AI
• that AI is only doing one thing: run a very simple tree
• the tree is just a sequence with 1 child. The sequence has a service running in tick with one print node. The child is a task with just one print node
• you are not setting tick rates or doing anything else anywhere else

[slow bobcat]

You are finishing the task as soon as you call move to. You should only connect it to success and finish task as success = true , and to fail and finish task as success = falss

[sacred stone]

Good morning.
So does that mean BTs will be deprecated and only available until UE7 in favor of the newer STs?
I have been working with BTs for some years and preferred them, though I am very open to learn STs too.

[crystal hatch]

yeah, it means we won't be maintaining BTs once UE6 goes out, and we'll probably cut the whole thing for UE7 (whenever that is). Having said that, I'll suggest to the team that we extract BTs into a separate plugin, so it's easier to share and maintain by the community, in the far future.

[chilly nebula]

What is the function in green on your graph? I would look into that

[chilly nebula]

Sadly there is no good way to do that atm, but its something we are planning on doing. Currently you have to make your tasks (or property binding functions) pull data in, or you could use events with payload to push data in.

[deft palm]

thank you very much

[misty wharf]

Btw you can use a blackboard for this, it just won't have the nice gui as you get in BT's :)

[royal mist]

state trees are great. Shooter AI go brrrr

I just used member variables on the character BP to track data needed, its a bit messy but it works well enough

[wise sluice]

Hello ! I took the opportunity to ask this :
Any plan to improve or replace the AI Perception System that seems pretty outdated compared to other AI systems?

[misty wharf]

Outdated how? 🤔

I'm not sure how that's the case, plus I think it's actually not that long ago when there were some updates to IAISightTargetInterface

[wise sluice]

The stimulus has very few info and event tag is a FName while we could benefit of a gameplay tag container for instance which is easier to use
We have to modify the engine to make it more interesting

[misty wharf]

Ah, yeah I suppose there are some small aspects like that which could be improved

I wonder how much could you improve by using a custom sense though 🤔 since I think you could have custom info in the sense's stimulus data

[wise sluice]

No any new custom sense can't add much since all stimulus store the same data in the end (See FAIStimulus)

You can add some inputs to your event but it will be limited to what you store in the stimulus

[misty wharf]

Oh I see, I've not looked at it in a while, I thought you could have custom data as well :)

[royal mist]

gameplay tags are sweet, Ive been using them lately and they are such a boon

state and status effects are so much easier to track with gt than a stack of bools lol

[wise sluice]

That's why i asked if there is any plan to improve this API
In any case, i'm personally modifying the engine for my need

[misty wharf]

Let's hope that Fortnite needs it, otherwise it might not lol

[wise sluice]

Would love to see some StateTree from Fortnite as template project

[royal mist]

yeah thatd be wild

[wise sluice]

Wild life, Npc and so on

[royal mist]

this is the current state tree for my shooter ai

[wise sluice]

I'm personally really interested in modular AI StateTree architecture (such building AI Archetype at runtime by adding/removing states)