gameplay-ai

[royal mist]

teal = utility selection state, blue = task state (mostly contains tasks), yellow = action state (mostly contains child states)

ooooo that sounds so cool

[misty wharf]

I think Siggi mentioned something about being able to inject subtrees maybe that's something that could be used?

[royal mist]

yeah eh

that does sound like its in the territory

[misty wharf]

I have a complicated BT-based system on some of my NPC's, where the NPC's have a set of tasks and each task has an associated BT... The NPC runs a "top level" BT and if it goes into the "do tasks" part of the BT, it uses the task-specific child tree

[royal mist]

sort of like a utility selection thing hey?

or how does the task get selected :O?

[misty wharf]

The tasks are always ran through but it could be selected via utility score also

[royal mist]

nice

[misty wharf]

(It generates them when the NPC spawns)

[royal mist]

oooooo

Im just cracking into state trees

but ive been a fan of utility ai system for some time

[misty wharf]

I also have a newer ST based system which does something similar, but based on a shared task queue where the NPC picks tasks to execute from :)

[royal mist]

thats cool

[deft palm]

Blackboard for state tree? Could you be more specific?

[misty wharf]

You can create a BB comp in the AI controller, and create custom ST tasks which read values from it and allow using them from the ST

[royal mist]

interesting

[deft palm]

Okay

[royal mist]

that seems cleaner than storing random vars on the character BP

[misty wharf]

Yeah it depends on what you're doing but it could be an option

[royal mist]

I have a bit of data on my character bp that would be nice to modularize

or whatever

[misty wharf]

You could in theory also use a custom state tree context to hold values, but it's a bit more involved

[royal mist]

hm yeah hey

[deft palm]

I think I will just store the variables in my aicontroller

[royal mist]

state trees are cool tho, I like the hybrid design

its like best of both worlds

utility state selection, and rule based

I was trying to mash something up between STRIPS/GOAP and utility ai

and i think state trees are similar to what I was trying to do with that

maybe not exact but ye

[final prism]

Does anyone know how the navmodifiers work internally? I'm trying to figure out which part of the navmesh generation they slot in to, but I've only been able to make guesses. Like if I mark something as a null area, it properly erodes the polygons, but it never seems to call the erode part of the navmesh generation? Or where does it re-triangulate the polygons when a modifier is added?

[slow bobcat]

when the nav modifier is added to the scene:

• its bounds are added to the nav octree
• the tiles affected are marked as dirty
• a rebuild is called for the dirty tiles
• new polys are created for said tiles
• nav poly flags (always for the area and sometimes for the poly) are changed

You can do a quick test to track all that down:

• in the level blueprint listen to some key input and spawn a nav modifier (you can just spawn an empty bp with a nav modifier component)
• set a breakpoint in FRecastTileGenerator::GenerateTile
• spawn the actor

[harsh storm]

When are you going to write a dang blog post Bruno? We need this stuff recorded.

[slow bobcat]

It's taking me months of skipping morning gym to finish an article about our AI, then I might prepare a talk about it.
Next is the blog but... takes so long...

[elfin smelt]

NavigationWizards.mp3

[harsh storm]

You ain't skippin' the gym. You're just gettin' swole with your brain muscle. 💪

[slow bobcat]

XD

At some point I will post all this in a blog and share my neat flow graphs about nav generation

such an easy system to map
(irony off)

[harsh storm]

See, that's why it takes you months. You put effort into graphs.

[slow bobcat]

that thing is my insurance. I can't rely on my memory

[final prism]

Hm this does not seem to be the case for navmodifier only navmeshes

[slow bobcat]

ummm what? let me do a quick test

[final prism]

Then again maybe UE is trying to bamboozle me I'll try running it with performance off..

[slow bobcat]

check setting a breakpoint in FRecastNavMeshGenerator::EnsureBuildCompletion

ah ok sorry I got things mixed up.
Set a breakpoint in ARecastNavMesh::OnNavMeshTilesUpdated and that will trigger. You can follow from there

you will end up in FNavigationDirtyAreasController::Tick

and this is the stack that happens before it due to pending octree updates (that's when the bounds of your modifier are added to the octree)

[Inlined] FNavigationDirtyAreasController::AddArea(const UE::Math::TBox<…> &, ENavigationDirtyFlag, const TFunction<…> &, const FNavigationDirtyElement *, const FName &) NavigationDirtyAreasController.cpp:135
FNavigationDataHandler::AddElementToNavOctree(const FNavigationDirtyElement &) NavigationDataHandler.cpp:237
FNavigationDataHandler::ProcessPendingOctreeUpdates() NavigationDataHandler.cpp:648
UNavigationSystemV1::Tick(float) NavigationSystem.cpp:16167```

[misty wharf]

Nice charts :D

And yeah it's pretty complicated. I managed to figure out how to create a custom navmesh

Basically in my game I needed to determine whether player placed objects are reachable by NPC's based on certain rules... I ended up doing a navigation query starting from a certain point and ending at the object. If it was possible to navigate, then it was reachable.

But this needed certain custom rules like what objects should be ignored for this query, which was separate from how they should impact actual navigation, so I needed to generate a custom navmesh for it :P

[final prism]

Okay yes it does call generate tile.. hmm my assumption was that navmesh modifiers don't regenerate the entire heightfield, but only modify the compact heightfield and re-tesselate from there

[slow bobcat]

That I'm not sure. Never looked into it

[final prism]

ahhhh I see what's going on, I think I may yet be correct! When it generates the tile after a nav modifier moved it sets "regenerate compressed layers" to false, so it doesn't do the whole navmesh process, it just generates the navigation data

My theory is that navmodifiers don't do any erosion, they just expand the convex hull by the radius of the agent

[slow bobcat]

That you know more than I do. I have no idea about the low level generation aside some stuff in the detour library. It's a pending topic for me

[mighty hornet]

This was the bulk of my overhead. It's a spline relative position call that I run once per frame for every pawn. And was calling it again whenever needed in my tasks. Now I am just grabbing the cached value off the pawn and cut my frame time down to ~200-300us for each AI...or about 90%. Huge. I still want to optimize further though.

Curious how much efficiency could be gained by moving everything to c++. Any insight?

[misty wharf]

In my experience there can be quite significant savings from converting BP ST or BT nodes into C++ if they are being called often enough

[slow bobcat]

Anything math related is way faster in c++

There are some videos about the topic

https://youtu.be/IhdobKXv_vY

https://youtu.be/3q-toBiBRzA

That last one is off because that's in editor, which you should never use to measure performance. Bp's in Package builds are much faster

[serene fern]

So having an issue which is when the player actor is moved elsewhere in the world the AI seems to on them moving still think they can see them when they cant any ideas what might be happening ?

[slow bobcat]

If you are following it and the move to is to the player, the move to task re-adjusts the path to it

[serene fern]

I found the issue I was having race condition

when I was trying to clear the sense target it seems in the same frame it was also setting it

adding a delay though for like 0.0 forces the clear of the sensed target to be nulled so all good

[misty wharf]

Tbh I almost always profile in editor lol

It probably isn't giving a good comparative time, but at least so far it does seem to tell me if my code is taking longer than it should even if it does have some editor overhead :D

[serene fern]

is there a way where I can force the Nav mesh to be a bit more into the centre away from the sides ??

at the moment AI is taking the path literally hugging the wall

only thing I need to keep in mind is theres doors that well it needs to get through so I can't really quash it down and is there a way to do it where if they are chasing the player they kind of ignore the edge being off limits is a way

[slow bobcat]

And just like that... I hate you. Hahaha yeah, it's true that for a general sense is ok. But very far from the reality of a build

[slow bobcat]

You can add nav modifiers with null to your hallway walls

[serene fern]

and then can I have it ignore them when chasing the player ?

bassically so the player can't hunker down near edge of the wall and it not be able to get them

[simple moon]

anyone have and tips/ a video recommendation to make my ai move? now it just stays at the same place and wiggels 😛 the world is procedural generated like minecraft, so it wont stop expanding aslong as you run 😛

is there a way to add navmeshbounds to the player controller? or something?:P

[mighty hornet]

Thanks. That's good to hear since a lot of my tasks are math based. AI needs to be as lean as possible since it's a racing game and I'm going for high fps.

[slow bobcat]

No that's gonna be trickier.

[serene fern]

arrr

[slow bobcat]

You could could try different approaches. You could have another nav aganet defined with a bigger radius. Use that to patrol/move and the green one you already have to attack/chase the player

[serene fern]

the other option is I make a system that has it follow points in the world and them points are places say in centre of cross parts

[serene fern]

I think I tried a while ago another navagent cause I think I wanted say some of my AI to be able to use vents where as others can't but I don't think I could ever get it working

[slow bobcat]

It's a but tricky because you need to pass the info to the move to request to choose the nav data you want to use, and that's only possible in c++ if I'm not mistaken

[serene fern]

my project is cpp so I don't mind doing that just be a cpp task I'd have to make

[slow bobcat]

Another option is to post-process the path to move it away from corners. Unreal has some functions for it. Also in c++

[serene fern]

hmm i might try the first one initially and see how I go

if that fails I'll attempt the second option you've mentioned

[robust pumice]

Hey everyone,
I'm currently working on simulating a realistic daily routine for NPCs in Unreal Engine (Blueprint only).
Each NPC has a Schedule array containing structs (NPC_ScheduleEntry) with fields:

Hour (Integer)

Activity (GameplayTag)

I get the current in-game time via Ultra Dynamic Sky → TimeOfDay, and store it as an integer CurrentHour (e.g. 900 = 9:00 AM).

What the logic is supposed to do:

Loop through the Schedule array (using ForEachLoop).

For each entry: check if Entry.Hour ≤ CurrentHour.

If true, save the entry as LastValidEntry.

After the loop, use LastValidEntry.Activity to set CurrentTask.

** The issue**

Even though CurrentHour = 900 and the only schedule entry has Hour = 900,
→ LastValidEntry ends up being None,
→ so CurrentTask is never set correctly.

I’ve confirmed that the schedule array has the correct entry and format. I suspect:

either the ForEachLoop condition isn’t working as intended,

or the assignment logic is firing incorrectly.

Maybe I should use ForEachLoopWithBreak instead?

Any ideas what might be causing this or how to fix it?

(Screenshot attached)

[misty wharf]

Where are you saving it to last valid entry? In your first screenshot, you have a set node for it, but nothing is connected to the input pin

I'm gonna guess that that's probably the problem :)

[chilly nebula]

A few games are already doing this 🙂 we added functionality recently to make this easier to do, take a look!
https://github.com/EpicGames/UnrealEngine/commit/d5b44e9281bdcebbbb9fc79f58259a7b1af633ab

[chilly nebula]

As already mentioned above, BP call overhead isnt very noticable except on code you call very frequently like on ticks. I recommend Ari's talk here that goes a bit into and measure the exact overhead.
https://www.youtube.com/watch?v=S2olUc9zcB8

[livid scroll]

Why is this always stuck here in this condition?

always in a state of fleeing?

[harsh storm]

Did you remember to call Finish Task?

[livid scroll]

Yes, i changed my task a bit

now it looks like this, what is this green arrow?

now its stuck between 2nd & 3rd

[harsh storm]

Time to bust out the visual logger

[livid scroll]

yeah good point, btw

look at this, a friend of mine told me to hook both the On Success & the On Fail to the Success (True)

I feel like this is wrong

but my friend is really much more advanced so I thought this is a good practice

[harsh storm]

I mean, I wouldn't say the task succeeded if the moving portion failed - but that's just me.

(And probably a whole host of other people 😛)

[livid scroll]

the person who told me this uses it in all his tasks

[harsh storm]

Don't mean it is right

When the task reports that it is a success, even though it actually isn't, it can screw up the actual logic of the BT

Sequence - execute each task left to right until the first one FAILS
Selector - execute each task left to right until the first one SUCCEEDS

So - by you saying that every result succeeds - you will always move to the next task in a sequence, even if you end up in an invalid state due to the move to failing.

https://dev.epicgames.com/documentation/en-us/unreal-engine/unreal-engine-behavior-tree-node-reference-composites

(I probably worded it weirdly, but the point remains)

[livid scroll]

what im trying to do is, make a chicken move in a more realistic manner, I was hoping of making this chicken have more complex logic, than just move from random radius to radius

[harsh storm]

You'd need to use EQS then

Gives you more control on how to do the pathing

[livid scroll]

like for example

1. give birth to other chickens
2. eat
3. sleep
   that's mostly it

[livid scroll]

i've heard of that

and yes I think its a good idea

[harsh storm]

This video is still relevant today
https://www.youtube.com/watch?v=iY1jnFvHgbE

Goes through the main AI tools

[livid scroll]

I saw this video btw

[harsh storm]

Watching and understanding are two different things

[livid scroll]

And I've read this documentation too, but still struggle to implement

[livid scroll]

exactly

[harsh storm]

Then you should've been able to tell your friend that he was wrong in his advice.

[livid scroll]

I dont undertand it until i actually implement it

[harsh storm]

Understanding is the key. Read/watch 100 times if you have to.

Baby steps.

[harsh storm]

Also for what it is worth - the far right sequence will never run.

[livid scroll]

I also have more problems than that 😂

my chicken doesnt look at realistic anyways even with the way it moves now

because of the animation blend space

and the model is just kinda OFF

I might need to first learn how to rig or something

but i've been on this for 1 year now

[harsh storm]

The chicken? Godspeed 🫡

[livid scroll]

My game is almost done in terms of everything else but the AIs

yes, and the bear 🐻 🐻‍❄️

which is the bad guy mostly

but I haven't set much logic into him, he just follows you around and gets in front of your face till you wanna break your monitor & keyboard with your head but thats pretty much it

[livid scroll]

yes at the moment it doesn't work indeed

[livid scroll]

do you think this will be easier to make with state trees?

somehow my brain is able to understand everything else & work with everything else just fine except for this AI programming related shit.

I hate AI programming the most out of anything in Unreal so far 😂 😭

[harsh storm]

Personally - I don't think STs are easier than BTs right now. But I may be in the minority on that. But I also don't think the tool is the problem.

[livid scroll]

they say STs are less complicated

[harsh storm]

Yeah, people like new & shiny things.

Here is my most recent gab about the two tools #gameplay-ai message

[livid scroll]

question about youtube videos, when following youtube videos that are not exactly always applicable like Ali Enzoheri's videos, there are over 20 videos in one of his playlists with 1 hour long videos each, he has his own assets, he often pulls his own little examples that have zero application to my problems, so there are so many videos made by him, not so much by other users, but besides that it's maybe even less than 5% of those things can be useful.

And most of the other videos on youtube made by popular UE youtbers are just the simple patrol - chase AI basic logic which they copied this from the official devs.UnrealEgine.com which is provided by UE documentation itself but they're not adding anything new unfortunately, most of the videos are kinda like this, and ytb is soooo limited in terms of AI programming😕

[livid scroll]

I'll give that a go tho, I think it now has more value to offer than I initially thought when I watched it few months ago

[harsh storm]

So uhhhh - what's the question? lol

[livid scroll]

oh I forgot to ask the question 😂 😭 😆

I guess no question, I guess thats it, just wondering how or why it is like that

seems like behavior trees have been around for so long too, almost as much as UE itself

kinda disapointing actually

thats it

🤔

[harsh storm]

BT has been around for at least a decade (at least in UE). As for why no one has done a good in-depth tutorial about it? Because the people who know how to use it are too busy working on games.

[slow bobcat]

Applicable to:

• gas for Multi player
• navigation
• path following and detour crowd control

[livid scroll]

seems like 90% of those videos on youtube are just showing you how to make a car but not how to make the engine, let alone help you understand how the whole thing actually works 😂

Seems like digital era snake oils salesmen some of them videos

[harsh storm]

I watched a little of Ali's stuff because a lot of people praised it - but early on, he makes some mistakes that can easily cause problems. So I attribute his popularity more in his teaching style than content. He does do more than the average person who does a one-off video on AI though. I'll give him that. Like, he actually show off how he approaches doing ranged combat and stuff. As well as mixing everything together. So there is at least some exposure.

[livid scroll]

or maybe its mostly about the views & likes not really even capable of teaching you how to tie your own shoelaces or smth

[harsh storm]

As for all the other AI videos that just show you how to chase the player for the 82934298457259th time... 🤷

[livid scroll]

yes I think Ali is a little bit more experienced than the Gorka's of youtube

[harsh storm]

Won't find someone altering the path points in a YT video - that's for sure.

I've been wanting to do an actual in-depth video/series, just...no time for the level of quality that I want.

I even have my notes saved somewhere on my computer.

Plus - the general audience of YT doesn't want actual solutions, because they can be quite time consuming to implement AND explain. They want to solve their problem here and now.

[slow bobcat]

If I ever start explaining stuff for the shake of giving back to the community, this is why I will never do videos. Good old GitHub tutorial with gifs and code you can copy and things I can correct if needed

I already did an attempt I never updated, so it's probably not accurate anymore. That was for a "simple" thing and took forever to post. Can't imagine posting actual complicated stuff

[gentle hamlet]

I realized despite tutorials saving you time in the moment, it totally breaks apart when you want to do something on your own

[slow bobcat]

Yeah, that's learning in a nutshell

[robust veldt]

It is a bug from nav modifier which does not update nav mesh when actor is attached.

[slow bobcat]

Attached to what? The player to the volume or what do you mean?

[robust veldt]

There is a spawner actor that spawn these collision actor at default transform and then attach snap to spawner. Nav modifier at default transform does not removed. It is reproduce on UE.

[slow bobcat]

Not sure I follow:

• a spawner spawns the nav bounds volume at a default location
• snaps the nav bounds volume is then move to snap the spawner location
• a nav modifier volume is not removed.
  But where's that nav modifier volume come from?

[robust veldt]

Sorry, Spawner spawn collision actor that have scene component as root, a box collision and a nav modifier component.
Then snap collision actor to spawner and observer that nav modifier does not clear at old location.

[slow bobcat]

Ah ok, so the nav modifier is moved but the nav is not re-built upon moving it?

[robust veldt]

It is still re-build at new location, but not at old location.

[slow bobcat]

Aaaah OK ok. That's strange. If you debug it, do the old tiles get marked as dirty?

[robust veldt]

As I debug, the nav modifier still report nav area at old location because child component not yet updated at that moment. I make a dirty fix by set timer next frame in UNavModifierComponent::OnTransformUpdated

In USceneComponent::PropagateTransformUpdate method, TransformUpdated is broadcast before UpdateChildTransforms that make nav modifier get wrong location of child component.

[slow bobcat]

Is the nav modifier a child actor or a nav modifier component?

[robust veldt]

it is nav modifier component.

[foggy sedge]

hiya!
im new to ai, been following a bunch of tutorials. im trying to do a simple third person ai that just sees my player when they are detected by the sight perception. i am running into an issue where the the player is nto being detected if they are behind some geometry that goes roughly to their waist.
i tried raising eye height of ai but that didn't really fix anything. any ideas? is the sight perception line tracing to the bottom of the player collider?

[misty wharf]

By default the trace goes to the actor's origin. The eye height is used as the starting point for the trace, but it doesn't impact the end point. To change the target, you need to implement IAISightTargetInterface on the actor being targeted https://zomgmoz.tv/unreal/AI-Perception/IAISightTargetInterface

[foggy sedge]

@misty wharf awesome thank you!

[livid scroll]

I think it will be a good idea

btw my chicken ai was kinda working before

but I decided to change it once again just so that I can kind of learn how different things work by playing around

I realized that I still have no clue how this fucking TargetActor is Set works

or what it means

its a boolean, I get that,but where is this logic at?

I haven't hooked up any logic to it

so how is my editor gonna favor this one in some cases if the chicken for example sees the player

& then I can also add a BTTask_flee

for example the blackboard right now never gets activated at all

I still dont really know how this works either

but it kind of gets the job done (almost) but I still need to learn how it works because Im gonna need to add more complicated stuff in the future

[solid wedge]

Difference between (default)AI Move To Vector vs Actor? Move to vector is convenient but updating a moving target made me switch to move to actors (empty actors existing simply for moving navigation goals). Any insights?

[royal mist]

state trees are so great! this robocop ai works so well now

Tho they can kind of be a pain to debug, even with the debug tools

[harsh storm]

It is a decorator. Pretty much, you will go into that branch IF the BB Key TargetActor is set (so not null). Otherwise, you won't. The logic for setting the TargetActor - that is somewhere else in your code. Typically it is handled by AI Perception. But that isn't the only way.

[deft palm]

Hi Siggi: I implemented my StateTree Task like this

USTRUCT()
struct FStateTreeSetGlobalResourceAndDepositBaseTaskInstanceData
{
GENERATED_BODY()

UPROPERTY(EditAnywhere, Category = Context)
TObjectPtr<AActor> Actor = nullptr;

using ResourceNodeRef = TObjectPtr<AAResourceNode>;
UPROPERTY(EditAnywhere, Category = Parameter)
TStateTreePropertyRef<ResourceNodeRef> ResourceNode;

using DepositBaseRef = TObjectPtr<ABaseBuilding>;
UPROPERTY(EditAnywhere, Category = Parameter)
TStateTreePropertyRef<DepositBaseRef> DepositBase;

};

in the task's EnterState:

auto ResourceNodePtrRef = InstanceData.ResourceNode.GetMutablePtr(Context);
*ResourceNodePtrRef = ResourceComponent->GetCurrentResourceNode();

auto DepositBasePtrRef = InstanceData.DepositBase.GetMutablePtr(Context);
*DepositBasePtrRef = ResourceComponent->GetDepositBaseTarget();

I got a compilation error:

StateTreePropertyBindings.h(969,57): Error : no member named 'IsValid' in 'UE::StateTree::PropertyRefHelpers::Validator<TObjectPtr<ABaseBuilding>>'

My question is: how to set and get the global parameter of a object reference in the task?

Thanks a lot

[misty wharf]

Ugh I thought I was being smart implementing GameplayTaskOwner on a Behavior Tree task, so that the BT task could auto-cancel any gameplay tasks launched from it... but of course the interface's function aren't getting called 🤔

Man, this is kinda annoying. Looks like in theory it works - if you use NewBTAITask in C++

So I have kind of elaborate logic that I needed for my AI, and I put it into a singular BT task, instead of splitting it into multiple tasks... because it would've become kinda noisy in the tree, and making choices in the BT tree isn't very nice either, but I can just easily do it in regular BP logic in the BT node

But because of that I need to do more bookkeeping of any AI tasks that I'm running from the node which is kinda annoying :P

Would've been great if it would've been able to track those easily which feels like the whole point of the BT task being a gameplay task owner, but I guess not

[slow bobcat]

if the bt task is the owner, doesn't it force you to have the task instanced?

[misty wharf]

Possibly yeah

It's currently implemented in BP's so it's gonna be instanced anyway :)

[chilly nebula]

What kind of type is ResourceNodeRef? btw if the only thing you are using your task for is to get an external type, its probably easier for you to just use binding functions.

Property refs are useful for modifying properties that are already defined in the state tree (in another task, state or global parameters).

[robust pumice]

Thank you thats it!👌

[livid scroll]

Ok so in that case my BB key perhaps didnt work because I dont have any function made to handle that in the AI perception BP?

oh wait but I got this one right here

IIrc I followed a tutorial for this, kind of a weird setup

I thought there would be a decorator blueprint where those things would need to be set up

or a blackboard blueprint or something

So all of these things here will be connected exclusive to the AI controller?

does that mean I can also set up "services" tpye of functions or whatever in the AI controller and it will work as if im creating a service

and how will I know when those things will FIRE up?

I can make a million functions & corresponding blackboard keys and they will never function or work if they're not somehow connected to anythin g

Ok as I see it this is hooked up to an actual event here

but these are the only type of variables that can be hooked up to actual events

[harsh storm]

There is a lot to unpack here

[livid scroll]

As I see it I get certain events based on the kind of actor components I've set up here?

[harsh storm]

Decorator is its own class. You can create custom decorators that test some condition if you want. But Epic provides some pretty useful ones out of the box. Such as making sure some BB key is valid. Which is what you're doing right now. Making sure TargetActor != nullptr pretty much (might be an IsValid() check under the hood, but its old code so not sure)

Service is also its own class. You use them primarily for the tick support. So you can do things while the node is running.

These two things are for the Behavior Tree itself. Not for an AI Controller or pawn or w/e else.

When you communicate to the BT from outside of it, you will mostly go through the AIController because it has access to the blackboard. Other things can obviously get the reference if desired - but it is game dependent.

AI Perception is something unrelated to the BT - but it is often used to update values in the BB.

So you could use w/e you want to communicate to the BB (doesn't have to be the AI Perception stuff)

[livid scroll]

Im double clicking on these blackboards, and nothing opens up, I think i've seen that in youtube where the creator is double clicking on the blue & takes him to a graph

[harsh storm]

Because those blue nodes are in C++

You didn't write them

They're provided to you by Epic

You need to have an IDE up (and maybe the debug symbols?)

If that even works for them

But if you make a BP decorator and double click on it - it'll take you to that.

[livid scroll]

Im supposed to have a BP decorator here am I not?

[harsh storm]

Why?

[livid scroll]

because if not then how else will this selector work?

[harsh storm]

By making sure the TargetActor is not set (so set to nullptr in the BB)

[livid scroll]

so is it set or is it not set?

[harsh storm]

It isn't. This is just setting up the BB

[livid scroll]

wtf does being set mean here 😂

[harsh storm]

If it has a value

Your logic for setting it is currently being handled in your AIController

[livid scroll]

so how can I set it?

that means my blackboard is set up wrong

and it will never work, right?

[harsh storm]

No

You can't set up a blackboard wrong

It just holds variables.

[lavish vault]

What is the difference between State Tree and State Tree AI in components?

[harsh storm]

State Tree is generic and State Tree AI gives you the AI Controller in the context.

[livid scroll]

ok so I assume this is the node that sets up my blackboard value

are you seeing a character or are you not seeing a character

technically any other actors are also characters, so perhaps I'd need to set up logic to confirm if it's a player chaaracter VS if it's another AI character

I left the SpottedActor variable unconnceted so perhaps this is forcing the 2nd one to be forcible UNSET

it needs to remove the value to be correct on the 2nd one

[harsh storm]

This is how the BT will flow:

The selector will first check the first sequence node and see that it has a decorator on it. Then it will run a method on the decorator that performs the test (I forget what the method is called). If that test returns true - it will then go into the sequence node and start executing the tasks left to right. If the sequence successfully executes each task, it reports that it was a success and then the BT will start the process over again. If one of the tasks fails, it stops right there and starts this whole process over again.

If the decorator returns false, then the selector will go to the next branch (your middle sequence node).

Rinse and repeat.

[livid scroll]

I followed a ytb tutorial for this too. So I didn't know 100% what I was doing

[harsh storm]

Probably because they didn't explain it. Or glossed over it very quickly.

[livid scroll]

isn't it gonna stay in the selector first and follow all the selected pathways?

[harsh storm]

Only if the decorator evaluates to true. And then it'll follow the sequence logic.

[livid scroll]

ok maybe not actually, because selector only selects one thing, and then moves on to business as usual

[harsh storm]

Once the sequence does its evaluation, the tree repeats the entire flow.

[livid scroll]

so what are some other types of blackboard conditional variable examples that can be set up with this?

[harsh storm]

Just look through the decorators

[livid scroll]

are the possibilities unlimited or they are kind of forced limited to what the events are in this blueprint

for example for move completed, I can make a float variable

[harsh storm]

Those events have nothing to do with the topic

[livid scroll]

for On Instigated any damage, I can make an integer variable, on Possesses Pawn Changed I can set up an actor variable, on Actor Begin Overlap -> actor variable again...

On Clicked --> boolean variable

etc.

[harsh storm]

I don't understand what you're trying to convey.

[livid scroll]

but the decorators have to be connected to these events

[harsh storm]

No they do not

The decorators have no knowledge about those events

[livid scroll]

can I set up these decorators outside of the AI controller?

[harsh storm]

There is no connection to be had

[livid scroll]

yes

I mean, blackboard variables can only be validated here, thats what i was asking

[harsh storm]

No

You can validate the BB variables anywhere

It is just a bag of variables

Grab one by name and test it

You just need access to the BB

The decorators are just there to be able to do tests.

So the sky is the limit on how you want to do tests

(Before Bruno comes in and ackutallys me - you can do more with decorators due to inherited functionality, but they're most often just used for test logic)

😛

You use one of the Epic provided ones if they fit your requirements. Otherwise, make your own.

[livid scroll]

Ok so in this episode of Myth Busters I've set up here on my first person character some keys, on the 2nd key its supposed to invalidate it, right? lets see how this magic works

[livid scroll]

Im pressing both but it's not really changing or affecting anything, perhaps blackboard chicken is invalidated, and I might need to cast it or something

[harsh storm]

The blackboard itself may be invalid

[livid scroll]

Ok Im pressing both keys

it's stuck on the Is Not Set

& waiting

Ok so for now I need to cast it

Im not sure how I should be doing these casts tbh

never worked with blackboards or chickens or whatever

I guess I'll do Get actor of class

It works but reallyyyy slow

I guess thats because it first has to finish the waiting time

which is very wrong

it seems like the wait is very unecessary or kind of breaks the natural flow of processes

[livid scroll]

perhaps I could be doing the wait through the Ai controller blueprint too, so that it can be later overrided by blackboard based variable

[lavish vault]

I'm following a tutorial on state trees, 2 years old, and when he creates override event -> Enter state, he gets a function. Whereas I just create a general event tick (in event graph). Which means I don't have the Return node with return value (succeeded etc.)

Where has this been moved?

this is his vs mine

I mean I could just use the state tree finish task node i guess, if that's the same. However I can't find an equivalent for the "Running" enum

[harsh storm]

Gimmie a bit - very busy. I'll recap and respond later tonight.

[harsh storm]

Wasn't "moved". It was changed entirely. They're now events instead of functions.

For it to return "running" - you just don't call finish task

[livid scroll]

One thing that I really hate about STs or BTs is that everything is inside of something else and its really hard to describe it to someone, it's not what you're used to when for your whole life for example you've been programmign in blueprint and then you come into BTs and they make it so in such a way that nothing is intuitive in the least intutive way possible

[harsh storm]

To be fair - you didn't know what "set the variable" meant. That isn't a BT problem.

[lavish vault]

Thanks for the explanation 🙂

[deft palm]

'''
using ResourceNodeRef = TObjectPtr<AAResourceNode>;
'''
It is a object reference to AAResourceNode.(AAResourceNode is a subclass of AActor)

And Yes I have declared a global parameter in the StateTree.

This is the global parameters

And Btw, I need to use the global parameters of ResourceNode and DepositTarget to determine where the StateTree goes ("FindAndSetTarget" or "CollectAndFill")

[harsh storm]

Do we have any stat commands for the state tree?

[foggy sedge]

another ai question...

the default ai sight collision channel is set to Visibility, but both capsule component and the character mesh are set to ignore trace responses on the visibility channel.

(using the third person template)

what is the "normal"/"common" way to have this setup?

[misty wharf]

Perception assumes no collision from trace == visible

So it should work with that setup

[foggy sedge]

omg, ofc, thanks!
yes now it makes perfect sense.

nice website btw :)!

[rare ruin]

Never seen one but there is stats for sure i saw some in Insight

[misty wharf]

Thanks :)

[serene fern]

is there a way for a recast nav thing to hold it's settings so for instance if I tick on draw poly edges and Enable Drawing ?

at the minute loading into different maps I have to toggle them two options on for the one recast nav thing

[slow bobcat]

I don't understand what you mean at all.
What are you trying to do exactly?

[serene fern]

so I have two recast in my world cause I have two nav agent types

showing navigation only shows one of then, for the other to show I have to enable following options

and the show poly edges

[slow bobcat]

Aaaah yeah. That's shit. I had that problem in a project and zi needed up doing a console command to show one or the other

[serene fern]

I was wondering is there a way for these recast to keep them settings so I don't have to keep going in to have all of them show when I turn on show navigation, it's fine if not

[slow bobcat]

I think I can dig the code I did for swapping the debug. But I'm not at home. Will take some time until I sit down on my pc again

[vast nest]

why is ai pawnsensing not working for me? it worked intially but now its just not working.
here's my code:

[misty wharf]

Hard to say, you need to debug it. Eg. add prints or breakpoints to help you see what is happening. In general, you should use AI Perception instead, Pawn Sensing is old and has been superceded by AI Perception

[misty wharf]

Man, I wish there was more information on how to best utilize gameplay and ai tasks 🤔

I've been debugging a bunch of issues in some of my NPC systems, mainly caused by the complex interactions between them and a bunch of other things

For example, there's an issue currently where an NPC has a "pickup item" AI Task, but due to the complex interactions with the game systems, there are some cases where another pickup task is started before the previous one gets finished or canceled (because it seems that BT tasks might not actually abort on the same tick)

but it kinda looks like with the right settings, the gameplay tasks component can actually handle this 🤔

[royal mist]

oh is pawnsensing old? wtf is ai perception and how do I use that?

ah weird

[misty wharf]

https://dev.epicgames.com/documentation/en-us/unreal-engine/ai-perception-in-unreal-engine

[deft palm]

I did some research and found that:

TStateTreePropertyRef<T>Parameter;
InstanceData.Parameter.GetMutablePtr(Context);

using the above code to set a global parameter of a StateTree in a custom task can succeed, Only if the parameter's type is one of the few basic types(int, bool, float, FText, FString).

I think this definitely needs to be improved.
Making accessing external variables unnecessary complex.

[chilly nebula]

I havent had time to test your case just yet, but did you try any of the other PropertyRef variants? I know property ref has been used for actor refs in the past, and even the code docs has an example

` * // Reference to Vector, TArray<FVector>, AActor*, TArray<AActor*>

• UPROPERTY(EditAnywhere, meta = (RefType = "/Script/CoreUObject.Vector, /Script/Engine.Actor", CanRefToArray))
• FStateTreePropertyRef RefToLocationLikeTypes;`

I recommend you take a look at STateTreePropertyRef.h

[final prism]

So I'm trying to wrap my head around the new navlink builder functionality, but I think I'm missing something with how the navmesh generation works..
The navlink builder seems to run on the navmesh tile generator, what I don't understand is how can it generate links that span multiple tiles this way? (Which you can easily see is the case if you look at the navmesh in any decently complex level with this navlink generation enabled)

[deft palm]

Hi Siggi

• Property ref allows to get a pointer to selected property in StateTree.
• The expected type of the reference should be set in "RefType" meta specifier.
• Meta specifiers for the type:
• • RefType = "<type>"

•    - Specifies a comma separated list of type of property to reference
  

•    - Supported types are: bool, byte, int32, int64, float, double, Name, String, Text, UObject pointers, and structs
  

[near condor]

For folks working with State Trees, would any of you be able/willing to share some documentation/tutorials that go over Considerations?

I have been using this document from Epic, but something about it is just not clicking with me and I would love some examples to look at: https://dev.epicgames.com/documentation/en-us/unreal-engine/state-tree-selectors-overview#considerationtypes

[deft palm]

but if the type is TObjectPtr<AActor> then it is not supported

[chilly nebula]

Yes, and did you have something not supported by those? An actor is a UObject

[chilly nebula]

Maybe its a compiler/syntax issue, did you try using AActor* instead?

[deft palm]

AActor* is Okay

[chilly nebula]

So what is the problem then? 🙂 Just use that. Or ideally if you can, use a weak pointer

[deft palm]

TWeakObjectPtr<AActor>?

[chilly nebula]

yes

[deft palm]

Oh it is also worked..

why TObjectPtr<AActor> failed?

[chilly nebula]

Can you tell me a little about what you find confusing or what issues you are running into? We are evaluating some changes here for 5.7

[chilly nebula]

Might just be a bug. Do you have the option to test in 5.6 ?

[deft palm]

not at the moment

I will use weak pointer

thank you very much

[chilly nebula]

np! Hope it all works out 🙂 just remember that weak pointers can become null if the actor you point at is destroyed, so plan your code accordingly

[deft palm]

okay sure

hello, How can I link my StateTree's Node to a linked asset like the picture below at runtime?

Thanks a lot~

[chilly nebula]

Currently you need to create your own C++ base class and add the function, but they are here in 5.6:
https://github.com/EpicGames/UnrealEngine/commit/d44c23e4d18aa65d8bd2d3841ca66eb243e0391e

[deft palm]

Thank you Siggi. you are Epic

[near condor]

Yes, I can try my best! I will walk you through the steps I went while trying to work with them and what I thought each step. If you want to jump straight to the question instead, check the bolded area.

1.) In the State Tree, I pressed the button for New Consideration, picked the base, saved the new consideration, and opened it.

Note: This felt straight-forward.

2.) Inside of the new State Tree Consideration Blueprint Base, I notice the EventGraph is empty, which tells me to next check the FUNCTIONS category to see if it is a blueprint where the functions are intended to be overridden. I see GetScore and GetDescription, and create overrides for them.

Note: After spending time with Unreal, this felt somewhat straight-forward as well.

3.) I check out the functions, starting with GetDescription. I see that it is as advertised, and just for putting a description. I then check GetScore. GetScore confuses me a little, but I keep seeing in the documentation that we are primarily working with 0 and 1, and as a float, so I assume it is like 0% to 100%. I create what I figure to be a simple test (see Figure 01).

Note: I suppose here I am confused on what is done with the Return Value in GetScore.

4.) In wondering what to do with the Return Value in GetScore, I decide the next step should be to add it via the Selection Utility for two Scenarios (see Figure 02).

Note: You can see how each Scenario is configured in Figure 03 (Scenario_A) and Figure 04 (Scenario_B).

5.) I run the test with the State Tree Debugger window open and take a sample to look at. I notice in Trigger Transitions, it outputs the warning "Could not trigger completion transition, jump back to root state.", where the intention was the have it go to TestWait.

Note: I assume this is not to do with Considerations, but probably the task, Debug Text, not returning as completed?

---

So what I am not exactly clear on, is the float intended to be sort of how weights are used? And if so, with both Scenario A and B are set to a Weight of 0.5f, how does the Weight get used after the Consideration is performed?

Thank you in advance!

[misty wharf]

bClaimRequiredResources = true;
AddRequiredResource(UKotAIResource_NPCAction::StaticClass());
ResourceOverlapPolicy = ETaskResourceOverlapPolicy::StartAtEnd;

Damn I think this just solved the annoying synchronization problems I was having with my NPC's

This makes it so that if there's an existing AI or GameplayTask running on the NPC with the NPCAction resource claimed, it just waits for it to finish/cancel before activating another - and it works with the latent node for gameplay tasks so you don't need to deal with it at all

Just need to use GameplayTask derivatives for latent NPC actions

[chilly nebula]

I recommend reading up on Utility AI before going further, if you havent already. Generally a state is scored by combining the score of multiple consideration (one works, but is generally less useful). For example scoring range or health of target to decide the final score of the utility of attacking.

Then the state selector simply selects the state with the highest score and runs that.

[deft palm]

Nicely done, it is simple

[misty wharf]

Yeah I knew the gameplay tasks component had a bunch of stuff in it but never looked at it in much detail other than seeing the comment that claims it has something to do with gameplay abilities lol

which it really doesn't

[near condor]

Gotcha, will do, thank you Siggi.

[chilly nebula]

There are multiple GDC talks and books on the subject, and probably a lot of YT videos too 🙂 My favorite authority on this is Dave Mark.

[deft palm]

What is the type of ResourceOverlapPolicy btw 😁

I wanna take a look at it 😁

[misty wharf]

ETaskResourceOverlapPolicy, this code is for UGameplayTask subclasses

[deft palm]

Okay. Thank you 😁

[near condor]

I will dig into those right now!

[misty wharf]

I was curious what exactly does an AITask do on top of a GameplayTask - it turns out, almost nothing

void UAITask::Activate()
{
    Super::Activate();

    if (OwnerController == nullptr)
    {
        AActor* OwnerActor = GetOwnerActor();
        if (OwnerActor)
        {
            OwnerController = GetAIControllerForActor(OwnerActor);
        }
        else
        {
            // ERROR!
        }
    }
}

// ERROR!

[final prism]

yeah its just a bit of added ai controller stuff, both gameplay tasks and ai tasks serve the same purpose: Tasks that can run async

[misty wharf]

Yeah

[slow bobcat]

I think that the limitation is not whether the nav link is more than one tile away but how are the tiles set in the binary tree.

[final prism]

hm not sure where a binary tree comes into play here, my understanding was that the tile generator only runs on data based within the boundaries of a single tile

Based on FRecastTileGenerator::BuildTileCacheLinks at least

Maybe there's a step that occurs after that I haven't found

[slow bobcat]

I'm super new to this feature and it's another team memeber the one that worked with it just Friday (he changed the code so we only generate nav links in tiles under special volumes placed by desigers), but looking at the code, looks to me that bGenerateLinks is enabled in every tile when using the tool within RecastNavGeneratior.
Then there's this step (image) where it seems to regenerate the CompressedLayers data if you want to generate links

So I guess the first time it will do the whole nav data, then only the tiles modified and, by checking the polyRef of the nav link end, it will also re-work the end-tile within GatherGeometry

[final prism]

Yes that's the flag you have to enable on the navmesh settings to enable link generation, its probably propagated to the tile gnerator

[slow bobcat]

probably in FRecastNavMeshGenerator::Init(). I know some stuff is passed to the active tile generator in that function, like number of worker threads.
Maybe?
Sorry if I'm just guessing here. I'm trying to help you and me both. You are asking some few deep questions I also have myself but didn't have the chance to check

My guess is that, since all the tiles are marked to generate tiles, nothing really needs to be passed to the tile generator. It only needs to read data from the tile to know if nav link generation has to be done on the Pending Tiles (PendingElement in the image)

Ah, seems that this is also part of the equation

[final prism]

Yes that is the part I have been talking about

[slow bobcat]

That's within FRecastTileGenerator::GenerateNavigationDataLayer

ah ok ok I see

[final prism]

Since its within the tile generator, it should only have access to tile specific data, not neighboring tiles

I think it marks up portal edges for later processing though, I'm currently looking through the code flow

[slow bobcat]

Yeah, looks like it just marks data like "hey, this nav link starts in this nav edge and ends here" for the linkBuilderConfig

An looking at it, it seems that this is where the tree limitation I was talking about comes into place (I know about it because I have suffered it). I thing the tiles need to be neighbours

Never mind. I think just marks a "this is where the nav link starts/ends" and that's it

Said limitation might not exist anymore (I suffered it in 4.27)

it would be awesome to know the difference XD

[final prism]

Between those types of configs? Jump over is probably akin to jumping over a bench or something and staying at the same height at the end of it

[final prism]

dtCreateNavmeshData is as far as I've gotten, I do think there must be a post process step to stitch together navlinks crossing tiles 🤔

[slow bobcat]

my only guess is that the nav link data generated is stored in the general data (outside the tile generator) once it's returned and, upon processing the other tile (the end point of the nav link) the data is "stiched". Educated guess of course

[final prism]

Yes its definitely integrated into the data it returns in the create navmesh data step, however I haven't seen anything in the navmesh generator that implies a step after doing so

[slow bobcat]

I don't think there's a step, just another TileGenerator for the End Tile. If you place a nav link in the level I think it's treated the same way a nav modifier is and both sides' tiles are marked as Dirty and tile generators are run for eache

then one tile generator will say "I'm the starting tile and this is the start edge", the other generator will say "I'm the end tile and this is the end edge"

And "somewhere" there's a common nav link identifier to grab the other edge when a nav link start/end is found during path finding

[final prism]

Why complex low level systems always gotta be unreadable 😢

[slow bobcat]

beats me. If you have access to any AI (chatgpt free daily tokens suffice) you can copy/paste code and get a bit of general help with what the code does. I have done it for detour copule times as a sanity check

[slow bobcat]

you can try this from a cheat

{
    None = 0,
    Default = 1,
    SomeOtherNav = 2
};

void ACustomBaseRecastNavMesh::EnableDebugDraw(bool Enabled)
{
#if !UE_BUILD_SHIPPING
    bEnableDrawing = Enabled ? 1 : 0;
#endif
}```

`void UXXXX::EnableDebugNavData(ENavDebugFlag NavDataIndex)
{
APlayerController* playerController = UGameplayStatics::GetPlayerController(GetWorld(), 0);
if (!IsValid(playerController))
{
return;
}

UNavigationSystemV1* currentNavSystem = FNavigationSystem::GetCurrent<UNavigationSystemV1>(playerController);
if (!IsValid(currentNavSystem))
{
    return;
}

if ((NavDataIndex == ENavDebugFlag::None &&  CurrentNavDebugFlag != ENavDebugFlag::None)
    || (NavDataIndex != ENavDebugFlag::None && CurrentNavDebugFlag == ENavDebugFlag::None))
{
    playerController->ConsoleCommand(FString::Printf(TEXT("Show Navigation")));
}

// this ensures we create the nav Render component without making it visible
currentNavSystem->VerifyNavigationRenderingComponents(false);

// This is the default recast nav data generated by our default nav agent type. We generate our own Recast object, derived from ARecast
if (ACustomRecastNavMesh* aiNavData =
    Cast<ACustomBaseRecastNavMesh>(currentNavSystem->GetNavDataForAgentName(TEXT("Default"))))
{
    aiNavData->EnableDebugDraw(NavDataIndex == ENavDebugFlag::Default);
    aiNavData->RenderingComp->SetVisibility(NavDataIndex == ENavDebugFlag::Default);
}

// this is the nav data type generated from our other nav agent config. It derives from our custom Recast Data (ARecast <- CustomRecast <- CustomDerived)
if (ACustomDerivedRecastNavMesh* otherNavData =
    Cast<ACustomDerivedRecastNavMesh>(currentNavSystem->GetNavDataForAgentName(TEXT("SomeOtherTypeName"))))
{
    otherNavData->EnableDebugDraw(NavDataIndex == ENavDebugFlag::SomeOtherNav);
    otherNavData->RenderingComp->SetVisibility(NavDataIndex == ENavDebugFlag::SomeOtherNav);
}

CurrentNavDebugFlag = NavDataIndex;

}`

this logic relies on you using different Recast Data generated per nav agent (which I guess is mandatory? I'm not sure, haven't been in a project with different recast objects in a long time)

[final prism]

okay I think I might have found it: dtNavMesh::addTile is called during EnsureBuildCompletion of the Navmesh Generator. I assume that gets called after all tiles have finished building.
There's a piece of code commented "connect with neighbor tiles" in there, that might be it

dtNavMesh::connectExtLinks specifically

I assume this is the reason we can't have links that span more than one tile, it just checks all neighboring tiles and connects the links if possible

[slow bobcat]

ooooh nice find. Ok so my theory was comletely wrong. Good thing you trhow some facts there. Thanks a lot for sharing all this

[final prism]

No worries, yeah I'm currently looking at finding a way to calculate the full size of tile cache layer areas, however these are generated per tile.. So I need to find a way to connect these over multiple tiles to calculate the true area size (in case they are connected over multiple tiles)

[slow bobcat]

What's the use case? Why do you need to know that?

[final prism]

Can't go into too much detail, but its similar to this feature described in mikkos blog around crouch areas (which was never implemented it seems) https://digestingduck.blogspot.com/2009/09/area-progress.html

You typically don't want small patches of "crouchiness" so you need to remove areas that are too small

[slow bobcat]

Aaaah I see

Sounds a very interesting problem

[final prism]

Yeah I think there are very few or no games that have areas both the player and npcs can crouch under (without being simple enough to just have a navlink)

[light ravine]

Anybody got any tips or pointers for how to handle AI being able to path in sublevels that are unloaded?

I am thinking, that say that we are loading/unloading parts of a level for performance, but still want the AI that was in that level to be able to get to the player into the currently loaded level

I assume the AI would always be spawned in the persistent layer, but how can I make it so that they don't fall through the ground and so that there is pathing still available

[chilly nebula]

If you need them to "path" and the level isnt loaded, the only way you can really do that is to build a static graph or use something like ZoneGraph.

Basically simulate an approximation, and ideally without spawning the full actor. Then when the player gets closer and the level is loaded, you "upgrade" it to a full actor that navigates on navmesh.

[light ravine]

I see. There might be a missunderstanding on my part about the frequency of sublevels needed within a map to where I am overcomplicating the AI logic and navmesh for this. I was thinking that essentially sectioning of smaller rooms into their own section, would allow for far better performance than just normal visibility culling, do you know anything about that?

[deft palm]

Hello guys, I got a question here just out of curiosity:
"Object Reference" in the image, Is it a TObjectPtr or a TWeakObjectPtr?

[misty wharf]

TObjectPtr

I don't know if it's possible to explicitly declare/reference weak pointers in blueprints

[deft palm]

Thanks a lot😁 . why is it not possible~~

[unreal briar]

Hey guys,
Any advice for a good architecture to create squad system for an RTS game?

[near condor]

@chilly nebula sorry for the ping but these Dave Mark GDC hour long each presentations have been excellent and are allowing me to pull together concepts from my other autonomous programming work for concepting AI in UE. Thank you again so much, these are exactly what I have been wanting for a long time, and makes sense I could not find them since they aren't on youtube to my knowledge!

[chilly nebula]

Im very happy to hear that, glad you are enjoying the talks! Did you watch the "Making a better Centaur" one?

[final prism]

huh I never heard of ZoneGraph until now, looks very interesting.. It looks like its been experimental since 22? Is it still being developed?

[chilly nebula]

Its not being actively developed, no.

[slow bobcat]

Can't recommend that talk enough too

[near condor]

I started it, but he quickly mentioned he recommended people watch Architecture Tricks Managing Behaviors one first, so I pivoted to it to cover that material before returning.

[restive imp]

Hello,

Beginner with AI here,
I'm trying to make my AI Move and shoot at the same time with StateTree

I use the same Task to do it with Event Start -> Move To Location
And Even Tick -> Shoot at target

But it seem's that when the AI is in MoveToLocationOrActor Node it lock AiLogic until he reach the destination point is their a way to bypass it ?
If i uncheck it in the Blueprint he just stay in place and shoot.

[deft palm]

Two tasks (any of them is asynchronous) in the same state execute concurrently

If you uncheck lock ailogic box

You could put the 2 tasks in 2 states and trigger the transition to shooting state when arrived

[deft palm]

Btw there’s an excellent explanation of the execution of state tree tasks in the sever

And I quote

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.

[deft palm]

Learned a lot from it

[restive imp]

Thank's, I've split the task in two as you said and running them at the same time seem's to have fix the problem 🙂

[hasty turret]

Hi everyone. I've been working on implementing Sekiro-style AI using State Trees for a while. State Trees work well for handling combos and their transitions, but I'm running into trouble when it comes to logic involving timers, counters, and interrupts. For example, I want the AI to do a turn-around attack when the player is behind them, or trigger a sweep attack after the player blocks successfully a certain number of times(and then maybe dodge back and idle for a few seconds). These kinds of reactive behaviors are hard to represent cleanly in a State Tree. Any advice or thoughts for handling this type of combat-heavy, condition-based AI?

[near flame]

Hello everyone.
I struggle recreating the bad AI system that i had and i don't know how can i switch between :Idle ,Move To Arena,Strafe,Move To Portal states
how is the best way to switch between them?
I had an enumerator with this states and used it in the old system ,but i don't know if this approach is good or no

The selector works but i have no idea how to implement the states below
and also when i tried to connect this one ,it made by default the idle state,and after finishing the monologue,it was stuck at root

[chilly nebula]

Counters and timers are best to setup as parameters either on a parent state that wont change (ie Combat) or global, and pass them in as ref (see discussion on TStateTreePropertyRef above) to a task to modify/refresh them. Then setup a condition to check their value.
"Check for number of blocks" can be done that way, with a task that resets the counter on a failed block/parry and another to increment.
Timers can be set as time stamps, so when you enter a certain state you save the current time, and in your condition you check if current time is higher than saved timestamp + whatever time you want to wait.

Interrupts can be done with a transition with a condition, but you might need to do some of that as a ticking transition. 5.6 is bringing new features for better async tasks that could help here.

[fathom sun]

Hello, I'm trying to make so that my AI patrols along a spline following the CurrentPatrolPointIndex. However, I want it to pick the closest point when it starts patrolling. E.g., if it leaves the combat, I want it to pick the closest one. If it finished moving towards one patrolling point, and tries to pick the next one, I do NOT want it to pick the closest one, but use the one in the CurrentPatrolPointIndex.

I have patrolling as a separate BT for ease of use. I thought about using something like a service or a decorator to check whether the BT has switched to something else, so that I can pick the closest point. However, all the ReceiveActivationAI, ReceiveExecutionStartAI etc seem to be called anytime I start executing the patrolling BT in general, so it looks like I can't achieve this using either service or a decorator.

How would you try to achieve something like this?

[near flame]

Is there any reason for AI not to go to this location even if i see that the wires are active and Portal location is valid?

I use similar logic to this one and it works fine

but this one doesnt work,why?

[slow bobcat]

do timers work in BTTasks? Asking just in case.
Does the node run?
If you print the location (debug draw) is it actually a valid location?

[near flame]

yeah they work,cuz this one works

[near flame]

ill try it

[slow bobcat]

Also check visual logger navigation and path following categories, let's see what the debug says

[near flame]

you mean the nav mesh bounds volume?

i hear the first time about path following categories and visual logger navigation

[slow bobcat]

open visual logger, hit record and play your logic, then check the log categories.
If you don't know what visual logger is, there are lots of info online about it

[crisp vessel]

cant figure out what's wrong, everything seems to work, judging by string prints, but enemyai actor doesn't go to "covers", it just stands still

[crisp vessel]

there is NavMesh btw

[misty wharf]

Ugh, why can nothing with AI ever be simple lol

It just occurred to me that I need to support resuming smart object behaviors... because if the game is saved while an NPC is using one, it needs to be able to continue what it was doing when loaded

[slow bobcat]

Is that blue sphere your locatiom to go?

[slow bobcat]

This is the kind of things we would force a designer to change hahaha

[harsh storm]

I was just about to say - "just uhhh...cancel it"

[misty wharf]

heh

[harsh storm]

I solve all my saving issues by just not saving 👍

[misty wharf]

Yeah I mean majority of my SO's aren't going to have a problem with that I think, but I have some which might

I already limit saving to certain points because of this kind of things, but it's going to be problematic to cancel it entirely for some 🤔

They would still need to keep track of what they canceled and resume it

[slow bobcat]

You can save the memory of a tree and restore it later. It causes issues with nodes that clean memory upo exit (we don't restore engine ones, just ours). If you track everything in the BB, it's even easier

[misty wharf]

I'm using custom gameplay behaviors insteado f BT's because some of the behaviors would probably be pretty annoying to model in BT's with all the logic and choices :)

But I think I can make it so that if an NPC tries to activate the SO again, the SO can detect from its own internal state that it needs to actually do something slightly different

it's kind of a faff I wish I didn't have to save but it's the kinda game where that can happen :P

[crisp vessel]

the blue sphere is nearest cover location for the actor

not player

btw

seems like hes moving

but very odd

[crisp vessel]

why doesnt he use other covers?

these

[fathom sun]

Hello, is there an easy to way to randomize one of the three tasks in a selector? I want a way to randomize my AI actions, like strafing, guarding, making a thrust or a slash, or any other type of attack randomly, and with a little bit of weights depending on the environment and the type of target it's facing

[harsh storm]

You need to build some kind of utility composite node. There isn't anything out of the box for BTs. But there is something for state trees.

[fathom sun]

The thing is that you can't create your custom composites since 5.4

[harsh storm]

There are also 3rd party solutions that have it already.

[misty wharf]

Iirc there's an example of creating a random or utility BT composite somewhere on the internets

[fathom sun]

For instance?

[fathom sun]

I have made that in 5.1, but now that I'm on 5.4 it doesn't compile

https://github.com/EpicGames/UnrealEngine/pull/12403

[misty wharf]

Classic Epic just breaking shit cos they forgot an API macro

[fathom sun]

And even rejecting the pull request to fix that

[misty wharf]

Probably just the bot closing PRs that nobody bothered to look at

[fathom sun]

But what 3rd party solutions are out there?

[fathom sun]

oh, so should I try to re-open it?

[harsh storm]

Ones that have a utility composite node or decorator.

[misty wharf]

Maybe, but it might just get closed by the bot after no human looks at it :D

hmm actually

    AIMODULE_API virtual void InitializeMemory(UBehaviorTreeComponent& OwnerComp, uint8* NodeMemory, EBTMemoryInit::Type InitType) const override;
    AIMODULE_API virtual void CleanupMemory(UBehaviorTreeComponent& OwnerComp, uint8* NodeMemory, EBTMemoryClear::Type CleanupType) const override;

this is what the functions look like in 5.5

maybe it was done?

[fathom sun]

I guess

Is it the time to upgrade to 5.5? 🤔

[misty wharf]

It's past .1 and .2 so it seemed fine to me so far :D

[fathom sun]

Alright, I'll try it out, thank you for letting me know that they have fixed it

[lyric flint]

He y there
I’m working on a UE5 project where NPCs walk on a spherical planet, using custom gravity to stick them to the surface. Orientation and gravity are functioning great, but I’m stuck on getting AI pathfinding to work all around the globe.
i can only generate nav-mesh at the very top of the sphere, everything else is considered too steep (or otherwise unwalkable).

Has anyone successfully wrapped a UE5 NavMesh fully around a sphere?

[misty wharf]

That sounds like an enormous amount of changes you'd need on the generator code and everything else 🤔

[lyric flint]

Damm, i jsut saw some plugins do it but they are really pricy

[misty wharf]

You could try your luck emailing the plugin author :)

[lyric flint]

Yeah, but is it really that imposible

[misty wharf]

Maybe there's some other solution to it, at least if the navigation isn't necessarily that complex

Eg. you could maybe generate points manually on the surface of the sphere, and navigate using those instead

[lyric flint]

Thanks, i might try something like that 🙂

but how can i make the ai move without using ai move to

the move to location or actor doesent seem to work

[misty wharf]

You'd probably have to build a custom solution to steer it towards the points

[lyric flint]

shit okay thank you

[chilly nebula]

This one was closed since it was already addressed in another CL 🙂 before creating a CL, always check the latest version of the codebase

https://github.com/EpicGames/UnrealEngine/commit/c3ea41d7624ec9afe294592f86f407155310775d

[fathom sun]

Oh yeah, I definitely should've

[slow bobcat]

Are you projecting that location onto the nav or passing it straight to the move to node?

[crisp vessel]

second

[lyric flint]

Can anybody help me, i dont udnerstand why this always fails to move?
https://blueprintue.com/blueprint/-9_pv_ea/

The map is on a sqhere. The youtube video said that it would work on a sqhere too. THis is the youtube video https://www.youtube.com/watch?v=TtEZrKM4Ue0

Thanks 🙂

[slow bobcat]

Then that's probably your problem. That pint is very far from the nav. Use the project point to nav node and pass the result point to the move to node.

[harsh storm]

Let's see if I can unravel all the goodies @chilly nebula has been loosely advertising 😎

[harsh storm]

Looks like a toggle that we can do to have a task be evaluated for transition or not

I wonder if this is what Siggi was referring to when he said we could make the ST fully event based?

[chilly nebula]

Its part of the puzzle yeah 🙂 also notice the "Any" dropdown there which is the same as the old default, but you can now also set the transition to only trigger when All tasks that are being considered are completed.

This + tick optimizations and scheduling, and the new task delegates and listeners (more coming for those in 5.7 too).

[harsh storm]

Yeah - I saw that too. Thought it was a little awkward of a placement, but not like I have any idea on where a better place for it would be. Or maybe there just needs to be like "Completion ....something" kind of text next to the dropdown. I just didn't mention it because you've already shown it.

Haven't looked at any of the new task delegates and listeners.

[chilly nebula]

UI will continue to be improved 🙂 the delegates are opening up some new patterns that are causing us to rethink things on a bigger scale. You can basically create task graphs by chaining tasks together now.

[harsh storm]

I'll see if I can find those task delegates and listeners. I was mainly just looking at the BP STT stuff.

[slow bobcat]

Siggi are you guys going by heart or following some paper/pattern for how/what you implement?

[harsh storm]

Found it

[chilly nebula]

A lot (most really) of what we do is driven by something internal or a licensee we are working closely with.
How State Tree started I don't have the full picture of, since that was originally Mikko's baby.. but I know performance and scale were at the heart of that, and driven by City Sample demo.

[harsh storm]

I appreciate you Siggi. Just sayin'.

[chilly nebula]

key thing about delegates is they are bound like properties. You setup listeners and dispatchers on your tasks, and then hook them up via the bindings UI.

[harsh storm]

Got it. So no payload is expected to be supported? Events should be used with those.

[chilly nebula]

Undecided yet, its something we are discussing internally. At least not in 5.6 🙂

[harsh storm]

So now, what we can do is something like:

State A

• Child B

Have A run a task that has some delegates. Then B can listen for those delegates on the enter state method. And respond in kind.

Man I hate Discords formatting sometimes. But you get the idea.

So A could set up all the data needed for its child states and then broadcast the delegate when it is done.

[chilly nebula]

yep! + you can setup so the delegate triggers a transition

[harsh storm]

Yeah - saw that too.

[chilly nebula]

What we are working on now (related to the payload discussion) is how to best deliver result of an async task that signals completion.. like a run EQS task. There are some issues with the current approach that we want to improve

[harsh storm]

I'd imagine y'all have until like November 😅

Can't imagine 5.6 would be the only version released this year.

So, some quick impressions - seems like this is a nice self-contained way to handle more event-y stuff. Because outside systems have no access to the delegate stuff it looks. So Events would still be preferred if outside systems need to communicate to the State Tree

[chilly nebula]

@misty wharf was it you who had issues with the RunEQS task and had to build your own? Sorry if Im mixing up who it was, from a few weeks ago.

We have some fixes in 5.6 for it (and the MoveTo) task, hopefully addressing the issue you had.
https://github.com/EpicGames/UnrealEngine/commit/c8c4f0baaae39b7b5664724911b9481405a38b42

[harsh storm]

It was me

[chilly nebula]

ah ok 🙂 even better then

[misty wharf]

What was the problem?

[harsh storm]

For some reason - the data just wasn't being transferred.

So the EQS result wasn't going to the move to vector

[chilly nebula]

A typical pattern we see is a ST task that registers to an external delegate and has an internal callback that broadcasts the task delegate. No need for events from outside

[harsh storm]

Thus causing the Move To to fail

However, when I wrote my own (using the BP EQS nodes) it worked perfectly fine.

[misty wharf]

Interesting, good to know if I happen to run into it

[chilly nebula]

yeah we've identified a few issues around that, some are fixed in 5.6 but at least one is related to the thing I mentioned above with needing a payload. The issue is if you have a task in a child state binding to a parameter of an async task in a parent state, it might prematurely copy the bound value and not update it in all cases when the async task finishes.

[harsh storm]

This is probably a better approach honestly. I was just thinking of it backwards 🤣

[chilly nebula]

This might be because of the BP itself or a Tick or State Enter or Exit firing and updating the bound value.

[harsh storm]

So, these three things - already 👌

[chilly nebula]

Duro Im curious if you try to set up your EQS+MoveTo again in 5.6 if it works or not. If it doesnt, I really want to get an internal repro so I can be sure we'll get that fixed.

[harsh storm]

The tick scheduling and optimizations aren't something that are within my purview as I've not had issues with those before.

[harsh storm]

Will do. I'll do it tonight after my day job.

[chilly nebula]

Basically anything that was ticking before will now automatically tick less, and only when it needs to 😉 + you can set lower freq ticks, or schedule the next tick directly. Like the delay task knows from the delay value when next to tick

[harsh storm]

I saw the wait for next tick or something just before I shutdown the editor.

But also - nice. Looks like we can employ the service pattern from BTs better now. (Controlling the tick rate was the big thing from it)

[chilly nebula]

happy to hear that 🙂 I expect we will adopt more and more of the useful patterns from BTs even though the VM and general structure stays the same.

[harsh storm]

Minus the BB, I think that handles most (if not all) of the ways to set up a BT. Because we could already use "decorators" as an enter state conditon. Replicating Services just needed to be able to control tick frequency. Because they're essentially just a task that ticks with some safeguards (which can be handled with the enter/exit state stuff). And to get the sequence logic, it is literally just having states transition to the next state on state success.

And selector logic is enter conditions and/or transitions on failure.

[chilly nebula]

yes, or you can chain tasks together to run in sequence.. but we need to improve our debugging tools to better support that 🙂

[harsh storm]

Can the tasks listen to delegates from other tasks on the same state?

[chilly nebula]

Dont forget you have state selectors in STs, but perhaps writing custom ones isnt as easy as it maybe could be

[chilly nebula]

yes

[harsh storm]

Well, I did forget. So there is that.

[harsh storm]

https://tenor.com/G6s2.gif

[chilly nebula]

I think the only requirement is they appear higher in the list.. the task order matters from the binding perspective.

[harsh storm]

That is fine by me and makes logical sense.

[chilly nebula]

There is nothing preventing you from setting up a whole task hierarchy inside one state now, ie a state that triggers 2 other tasks, and those trigger X many and so on and on...

Only downside Im aware of atm is the debugging tools dont show this very well

[harsh storm]

That would be hell. But nice that it is supported. Does this mean that y'all guarantee (at a minimum) that the Enter State function will be called sequentially?

[chilly nebula]

<@&213101288538374145>

[chilly nebula]

I dont think there is anything currently that sets that up, but that is a good point. However with the bindings going through property binding it should work, unless you have some init dependencies between the tasks?

[harsh storm]

I'm just thinking of a situation where, let's say Task A has a delegate and Task B has a listener. Task B needs to, on Enter State, set its callback method with that Receive Callback (or w/e the BP node was named. Only looked briefly) to some function.

What happens if Task A broadcasts the delegate on like the 3rd node call in the enter state function. IE - Enter State -> Do something -> Do something -> Broadcast.

As I type this out, this has nothing to do with what I actually brought up with sequential Enter State calls. So ignore I said that part 🤣 (because it would still fail)

The issue being - task B hasn't set up the callback function yet, but task A already triggered the delegate.

Mind you - I don't know if the listener property has a way to bind to a function in the property bindings part. I only had time for like 5 minutes of snoopin' around because I had to get back to work.

[chilly nebula]

I'll have to check if this can happen, but I think its unlikely that an external async callback will fire before the rest of the tasks can run their Enter State. If so, I think the processing order will take care of that and not process the task delegate broadcast until after all the OnEnter have run.

[harsh storm]

That is one of the race conditions that come to mind with the sequential task setup you described.

[signal dock]

Hi, new to State trees and wanted to ask what’s the best approach to handle two different enemy types (range/melee with distinct tasks for each). Does a single AI controller with two states trees make the most sense? If so would you override the trees or just destroy the ST component and re-add that component with a different set on the controller when possessing? Or just one big state tree, and inside each task check what enemy type it is and execute the fitting attack?

[deft palm]

I think you could watch the video Siggi recommend

Im very happy to hear that, glad you are enjoying the talks! Did you watch the "Making a better Centaur" one?

[signal dock]

..?

[harsh storm]

Did my tests - works now. Only as a sibling state though. Which is perfectly fine. Still doesn't really work all that well in the situation you described about child state binding to an async task param in the parent. Thanks for looking into this though! 🥳

[deft palm]

https://www.gdcvault.com/play/1021848/Building-a-Better-Centaur-AI

Please watch this video

[slow bobcat]

For heat you need to do you don't need infinite axis utility AI, which is what that video explains (or any kind of utility AI for that matter).
A common approach is to have a melee bt (or st in your case) and a ranged bt/st. Then each type of enemy uses the appropriate one.
Within each tree there will be concepts like "approach location/target", "attack" and others.
Inside said actions you can then plug in sub trees (linked trees in the case of st's) and have custom behaviors for each unit sub-type. Ie: a sniper ranged enemy will behave differently than a shotgun ranged enemy. When looking for a shooting location, the sniper will move far and behind cover while the shotgun will move closer to the target. That would be the "approach location" part.
When attacking, the sniper might aim, shoot, reload, repeat and the shit gun might just aim, empty the shotgun (2 bullets) and reload, being the reload much longer

The trees will look the same within a type, but the execution of the parts will be different and held in a different tree

That's a common approach

You could even have one general tree for all, but usually ranged look for cover points, try to pisition themselves out of the way of other ranged enemies etc. Their bt's are usually a bit deeper than melee

[chilly nebula]

btw I followed up on your race condition scenario, and it can indeed happen currently.

[robust pumice]

Hi, i'm currently trying to build a realistic daily routine system using Behavior Trees, where the NPC walks to a bench and sits down.

Setup:

A Blackboard Key (BankLocation, type: Vector) is correctly set at runtime.

The NPC has a Sequence in the Behavior Tree that contains:

MoveTo → target: BankLocation

BTTask_PlayInteractionAnimation → should play a sitting animation

Problem:

The NPC walks toward the bench but stops shortly before reaching it.

What could be the issue?

[ocean wren]

Use smart objects for that

[slow bobcat]

I don't see the stop you mention.
What I see:

• approaches the bench until the capsule touches it
• plays the animation (which for some reason is making the AI go away)

[chilly nebula]

yeah that makes sense 🙂 if you trigger a transition to a sibling/child state with the delegate the result data will be updated once you enter the new state. If you have a task in the same or child state registering to a delegate on a task in parent state, then the result data wont be updated in all cases when the delegate triggers listener code in the child task.

[near flame]

@slow bobcat i checked with the visual logger and here is the message it wrote

[slow bobcat]

I hope you have some intern doing UML graphs for documentation about this flows. Us peasants will consume them eagerly

[slow bobcat]

Yeah. It can't project the point to the nav mesh. It's probably too far or not so far but you are not passing a good projection extent

[near flame]

its because of the collision or i need to rise it a bit ?

[chilly nebula]

I see this as a bug that we need to fix, not something that users should learn 🙂

[slow bobcat]

your door is affecting the nav and creating a hole in it. The point is within said hole and you are not using any extent (allegedly) to project that onto the nav.
Project the point onto the nav and use the result in the move to node.
Or you can set the bCanAffectNavigation flag to false on the door (might cause other issues like AI's getting stuck when the path goes throuhg it)

[slow bobcat]

ah fair enough. Missed the bug part (catching up with the Durox conv)

[near flame]

yeah i succeeded,just added a billboard component on the teleport actor and used it as the location

Thanks for introducing me to a new tool

[slow bobcat]

the bildboard component is not going to work in a build. That's an editor only component. I suggest you use a scene component

[signal dock]

Thanks for the suggestions. I tried setting up two ST’s in the same AI controller but can’t seem to get it to work- any idea if it’s not something that can be done with BP?
Edit: Yeah, looked it up a bit and unfortunately seems like you can’t set a state trees on runtime.
Found this, I’ll give it a go:

https://github.com/gamedevkirk/UnrealEngine-ExtendedStateTreeComponent

[slow bobcat]

Rather than setting 2 ST's in the AI controller, you can just set the one based on the enemy type, make it more data driven.
You could do this on Posses before running anything else I believe.
The StateTreeRef is protected in the State Tree Component.
You could have a child of the component with a SetStateTreeRef being called OnPosses before calling the Super:: part of it that callls InitializeComponent() on it

hahah ok, just opened your link. Yep, exactly that

[signal dock]

Haha yeah thanks for the confirmation

Really wondering why Epic isn't providing a native implementation of it

But oh well

[slow bobcat]

Because it's very dependant on how you initialize things. If you were to swap the tree at some point, you will need to call the init functions again.
One thing you can do is just have linked trees in your main tree and use those depending on the enemy type

I think that's the idea in general: rely on linked trees

In our game we have the State Tree Component running several trees because parallel tree didn't exist and linked trees where a bit messy.

[signal dock]

I wanted to use linked trees but I just couldn't find solid documentation in the engine for it

[slow bobcat]

the code is your documentation

[signal dock]

well I'm trying BP first when possible.. and going to the online UE documentation to look things up

But realistically lots of times there jut aren't any

[deft palm]

I think this code could be implemented by blueprint

.

I also implemented the similar code. I think I was just coping the blueprint version of it

I wasn’t sure. It was weeks ago

[chilly nebula]

We do, in 5.6 🙂

[slow bobcat]

Is it auto-handled? What I mean is: does it Initialize the tree etc?

[chilly nebula]

One of the CLs:
https://github.com/EpicGames/UnrealEngine/commit/f6519f858472cd31ebea7c2ada2a332be58150cd

void UStateTreeComponent::SetStateTree(UStateTree* InStateTree)
void UStateTreeComponent::SetStateTreeReference(FStateTreeReference InStateTreeReference)

[signal dock]

Hi look at that, I just got a 5.6 preview notification 😄

Lets fire it up and see if these sweet ST setters are there

[chilly nebula]

😉 yup Preview was just released, enjoy!

[vestal briar]

I was wondering if it is bad practice to have a loop in a state tree where states are selected by tags that are set in a selector state within the loop.

I could use the Selection Utilities here, but I specifically want those states to be used based on cooldowns, timers, distance etc.

[harsh storm]

They are. When I originally wrote my own (all you had to really do was inherit from the component and then make a public function), I asked myself the same question. Then looked at ue5 main and saw that it was there. This was some months ago.

[harsh storm]

And I followed up on the tick frequency thing. I was expecting it to be settable per task - but it looks like C++ only? Because I didn't see way to alter it in BP.

Unless I'm more blind than Ray Charles.

[chilly nebula]

There is a setting for the tick behavior (scheduled vs something else).. I cant remember the exact name or where that setting is, maybe it was on the state itself? I can check later today once Im out of meetings

[chilly nebula]

btw the 5.6 roadmap is up 🙂
https://portal.productboard.com/epicgames/1-unreal-engine-public-roadmap/tabs/124-unreal-engine-5-6

[harsh storm]

Yup. We've been talking about it in #programmer-hangout

Looks like my exploratory stuff hit all the stuff that is coming for STs

And the main focus, gameplay wise, was ST and Mass for this release.

[chilly nebula]

I dont see a problem with that 🙂 but since you need to set the tag (presumably in a task), is there a reason you aren't using transitions instead?

[chilly nebula]

Mover 2.0 got a lot of improvements too 🙂 + gameplay cameras and other stuff

[harsh storm]

Yeah - but 8 out of the 11 cards were Mass/ST related.

That's what I meant.

[chilly nebula]

Doesnt mean there was no work done 🙂

[harsh storm]

Oh, definitely not. I'm not including stuff like the async streaming stuff for world partition or the zen improvements or the motion design stuff or anything like that. I was talking strictly about the gameplay framework section.

[slow bobcat]

This got me excited. Maybe next game...

And all the new stuff for level streaming and data layers looks awesome too

[harsh storm]

Yeah - maybe this is what is needed to get the gaming community to stop calling UE the stutter engine 😅

But it won't be realized for awhile. Because I think games are still mainly releasing on like 5.3 and lower.

[vestal briar]

Using transitions instead seems to work well and makes the inner loop look cleaner

[slow bobcat]

Tbh The stutter issue is on the dev side more than the engine.... That said, the engine doesn't make it easy to improve it

[harsh storm]

Well, yeah - but that doesn't mean anything to the end user.

To them - because so many UE games are plagued with it, it is the fault of the engine.

[slow bobcat]

Fucking gamers. They ruined gaming. Games should be done and then kept within the company server XD

[harsh storm]

You should link the talks to be honest

If they've proven to be so useful

Then we can bother the mods to pin it

[near condor]

@chilly nebula I finished these three talks

• https://www.gdcvault.com/play/1012410/Improving-AI-Decision-Modeling-Through

• https://gdcvault.com/play/1018040/Architecture-Tricks-Managing-Behaviors-in

• https://www.gdcvault.com/play/1021848/Building-a-Better-Centaur-AI

I also took notes, about 16,528 words with some drawings haha

[harsh storm]

Wrap them in <> to not have the embed pics

[near condor]

These really filled in the gaps for what I was missing. They also allowed me to tie together a lot of my current systems, at least conceptually, since I was building with the intent of modularity from the start.

So thank you

[harsh storm]

Helpful AI Talks:

https://www.gdcvault.com/play/1012410/Improving-AI-Decision-Modeling-Through

https://gdcvault.com/play/1018040/Architecture-Tricks-Managing-Behaviors-in

https://www.gdcvault.com/play/1021848/Building-a-Better-Centaur-AI

[near condor]

and thank you Duroxxigar for the tips there

[harsh storm]

<@&213101288538374145> Can we get these talks pinned?

[near condor]

honestly? yeah. AI is a ham beast, and I feel as though it should be required learning

[chilly nebula]

Happy to hear you liked them 🙂 if you want to go deeper, Dave Mark also wrote a book on the subject.

For State Tree we have a utility based selector, and are looking into some improvements for 5.7 so I'm looking forward to your feedback if you decide to play around with that.

[crisp vessel]

how to open this window in Learning Agents Interactor? I dont have it

nwm, I found it in "window"

it was hidden by default for some reason

[crisp vessel]

where can I find good q-learning tutorial for ue5.4?

grok, chatgpt and gemini have very strange info about learning agents plugin, it doesn't stick with reality

[chilly nebula]

You'll probably have more luck in #generative-ai

[hallow compass]

idk if this would be interesting to pin for new state tree users
notes of the basics of state trees with a list of other talks about it (most of them comes from the 2024 unreal fest)
https://notes.hzfishy.fr/Unreal-Engine/AI/State-Tree

[restive imp]

Hello , i'm trying to have my AI to move Physic Item to hide behind it what would be the best way to do it ? Put Tag on mesh able to move ? Overlapping Event with Collision on the AI ? any idea ?

[granite path]

Hello Ive been playing with AI and UE5. Well Im thinking that AI is actually the one playing with me. :/

[slow bobcat]

do you mean like opening a door and hiding behind or flip a table and use it as cover?

[wise sluice]

Thanks a lot!

[warped solstice]

i think this is the best channel for this question

how did you guys resolve problem of dealing dmg through walls? i mean case like melee AI attacking with long sword/hand whatever, how do you check if the dmg should be dealt or not?
trace from hand to dmg place to check if there is wall in the middle is the easiest solution, but i dont belive its good or safe, and may cretae more problems than solutions, so what other options did you try?

[misty wharf]

I'm assuming the reason for this in the first place is that you're using some kind of overlap check which just ends up going on the other side of a wall because of its size -- to me it seems you could do a linetrace from the actor to the point that was hit, and check if something is between them. This seems like it oughta work, although I've not faced this specific problem myself.

[shadow furnace]

In my opinion there is not really a good simple solution to this if the thing you trace from can be inside of the geometry that you consider blocking

My dumb idea would be to just trace from inside of the movement capsule against things that block movement in order to have a better chance of starting the trace above a surface rather than inside of it

[restive imp]

Flip a table en use it as a cover

[slow bobcat]

I woiuld look into smart object, they are designed for that kind of stuff. You can claim the table, play the anim, reposition the AI etc

[slow bobcat]

you could always keep overlap for damage in one collider and have another at the same location with the same geometry (let's say a capsule) that does physics collision block against walls, furniture etc (basically anything in the Static channel). The sword will not pass the wall. You can only enable the physics collison while the AI is doing the attack animation and disable it later (or as soon as it applies damage)

[warped solstice]

thanks for ideas, will check them again

[chilly nebula]

+1 on what people said above 🙂 Last I dealt with this problem was exactly this, a trace from "eye level" location within collider capsule to impact point to verify "reachability" of the point being hit.

[near condor]

I absolutely will. And I will see about getting that book, he seems to be a good teacher.

[near condor]

I could be testing this wrong... But I am confused.
Version UE 5.5.4

"I do not sense the Target." behavior = Try Select Children With Highest Utility
Screenshot indicates neither Idle nor Patrol have Considerations added to them, but the weight is adjusted for each.

I imagined that If:

  Idle is set at a lower weight (0.9f), without a Consideration applied, that it would treat it like 1.0f*0.9f=0.9f
  Patrol is set at a higher weight (1.0f), without a Consideration applied, that it would treat it like 1.0f*1.0f=1.0f

So by that logic, I would have thought it would default to selecting Patrol, never Idle.
However, it always picks Idle. Does this mean the "default" without a consideration is 0.0f, so then weights are multiplied against 0.0f, essentially zeroing them out, and since both would be Zero, and in the case of a tie, Try Select Children With Highest Utility would then select Idle?

I added a Consideration to both. A simple one that always outputs "1.0f".

I kept the weighting I had set, and now it selects Patrol every time.

[chilly nebula]

Btw always look at the debugger, it should give you insight into why something is selected.

My guess is there might be bug where this doesn't work without a consideration, since they are very central to how the scoring works.
Can you tell me a bit what you were trying to do here, and why you thought the utility would work without a consideration? If you only have a static weight, it will always select the same path.

btw. if you just want a weighted random, there is a separate state selector for that.

[earnest musk]

Hi all, small question regarding StateTree - what is the current status of evaluators? AFAIK they were to be deprecated in favor of global tasks, but they are still in.

[chilly nebula]

They are still there, we cant really remove them without breaking existing trees. They are becoming pretty much the same thing now.
Are there any cases where you prefer evaluators over global tasks?

[earnest musk]

Nope, just wanted to clear this thing 🙂

[serene fern]

So i'm having odd issue where a AI Move TO node is aborting but I can't work out way

the place it's trying to go to is valid

so i think I've found issue it seem to be choosing a non valid Smart object

[coarse smelt]

Hello, I've been using EQS queries in my BehaviorTree so that my Ai characters can find Actors of particular classes. Howver, I now have 20+ classes that my Ai might need to find and I feel like there must be a better solution. I want to pass an Actor class variable to the Ai character and have the character go to the nearest instance of that class. I've seen suggestions where I write a function in the character class to search for the class, and then (if I need to do more Ai stuff with it) pass it to the EQS as an EQS Query Context. Is this the easiest way to do what I'm trying to do? And a bonus question: is there a reason EQS Queries don't have an Actor input parameter? This feels like it'd be useful!

[valid basin]

HI guys, Im new to unreal. New to state tree. Im struggling with a simple setup.

Idle: Move to Random Location > Wait for 1 sec > loop
Chase: Move to Player when distance to player is < 250 unit
Back to Idle: when distance to player is > 250 unit

From my animation gif has shown. When player is <250 unit. The AI does chase, But if I move out of 250 unit before the AI get to the player. It gets back to patrol as expected. But when it get to the "wait" state, it'll try to get to the player again eventhough the player is out of its trigger range. It is like it hasn't been able to reach to the destiny and it keeps trying during the wait state. I have no idea why.

[heady silo]

I did a search in here and haven't found this yet, so I'll ask here:

I'm trying to come up with a generic function that can run for an AI character to locate objects in a certain radius that updates frequently and has a memory. I want to take those locations and put them in arrays for AIPerception to use as reference, and pipe these into data that BTTasks can use. In addition, I want to use the same generic function to highlight objects near the player that they can use in the world.

Let's say our NPC is currently unarmed and needs to find a weapon. We run this check to figure out what is nearby to use, and how to use it.
Say there's a knife and a gun. The knife is closer to hand than the gun, but the gun fits the more immediate need of a ranged opponent. But the ranged opponent is close enough that the knife can be grabbed on the way to the enemy to attack.

So we need to store some data-

The nearby objects as ObjectsInSearchArea(array)
Distance to each object from our starting position as DistanceToObject(array)
Types of objects as defined in a weapon class like BP_Weapon_Pickup

and reference these in the Task executed in the tree.

So I want to compare the generated list to what is seen in AIPerceptionSight() and get the closest object to move to it. I guess if I was to expand it to have more "decision making" I'd need to get the enemy's location via perception or their last known location and set distance values such that if the unit is close enough to attack quickly in melee to prioritize the knife, and further than that distance to priortize the weapon and getting to cover. Does any of this code even look close or am I way off?

I guess maybe this would make more sense in a GOAP system but I'm trying to reduce my learning overload by concentrating purely on what I can do with BTs for now

unless StateTree would be a better fit for functionality like this?

//revised

(ignore the tick usage, I will probably try and make this into a BT Service)

[vestal briar]

You could use gameplay tags on the actors for an EQS query rather than checking for specific classes.

[vestal briar]

The Debugger window in the State Tree editor is very useful for tracking down why you are transitioning unexpectedly to the different state. It'll show you the values in question and why it transitioned to your Chase state unexpectedly.

[valid basin]

It is still not making sense. it is clearly in the wait state. and the wait state has only 1 delay task. why is it moving towards the player.

[obtuse igloo]

Hey all!
I'm messing with Smart Objects, and I was wondering how I could get my AI to stop trying to move to and use a Smart Object after it's slots have been disabled? My AI is trying to move to and eat an item, but If the player picks the item up (it gets attached to the player's hands), I'd like the AI to stop trying to go to the object and abort it's task and un-occupying the slot. Any ideas appreciated ^_^!

[slow bobcat]

You can write your own eqs generator that looks for actors of class X within radius (or whatever you need). The generator can use a Query Context that will provide the class type. That query context can return the actor class you passed to the AI character.

[slow bobcat]

That said, this is a much better approach. G.tags for the win. You can make your generator based in g.tags

[slow bobcat]

while there are many approaches, your requirements would fit in an Utility System where you score the weapons to pick one

[heady silo]

Apparently Utility AI is a part of StateTree as of 5.5. Some more learning is in order

[slow bobcat]

That would be more to weight decision making. Like "go pick a weapon or run away" rather than choosing what to pick.
I don't know in BP's but in code you could funnel all this through EQS, have your own scoring system etc. Then run the EQS when needed and get your results already ordered by scores. Then pick the best one from there

[heady silo]

So could I just do something like sensing the class of the nearby weapons, and then picking the highest class with the shortest distance?

[slow bobcat]

that's what an EQS would do yeah.

• A generator will produced a raw set of data. ie: all the weapons within 25m (our give you all the available guns to pick in the game and a test will discard the ones further than X meters)
• then a test will score them based on distance (the closest the higher) <- this the engine already has
• another test will score them by how dangerous they are (a gun is more dangerous than a knife) <- this you will need to create by reading the Dangerous value from your weapon

When the eqs is done, it will give you back an array of weapons (items in EQS lingo) ordered by their score

[crisp vessel]

why I cant use attack function from enemy blueprint?

and how to use conditions in behaviour tree?

[misty wharf]

Behavior Trees can only use nodes created for behavior trees. If you want to use a function from your actor, you need to create a new Behavior Tree Task which wraps the call of the function.

For conditions, you would typically use decorators, but the way all this works in BT's is not quite the same as using if/else in BP's. If you don't know how this works, it would probably be good to read the docs on how Behavior Trees evaluate.

[chilly nebula]

Hey @harsh storm sorry for the late reply on this 🙂 If you change the "Scheduled Tick Policy" on the Schema settings on the tree to "Allowed", the custom tick rate settings on state will appear.

Its per state and affects child states as well.

[harsh storm]

That was a lot of meetings.

[near condor]

Apologies for the delayed reply! I was trying to toy with it more and some work came up. I was more so just trying to understand how it is selecting states, with and without considerations.

For why I thought it would work without a consideration, I assumed that the weight value is being supplemented as "chance" even without a consideration, and assumed that without a consideration, all default base scores were a value of 1.0f

However, it appears the "default" of the setup without a consideration is to be treated as though it were a zero, as weight holds no impact without a consideration as you mentioned.

---

Unrelated to the weighting, I was also noticing that with a consideration, they seem to not really accept Input variables?
But only sort of?
They can take in the Actor/Controller/Global Params, but if you create params on the State itself that "set" variables within the scoring method, it seems as though it just straight up breaks the Consideration.

Example: in the screenshot, I have the Input Var STC_TestBool, and the default value is true. Having that variable in the first place setup with a binding in the state tree is problematic, as the state tree seems to skip over the state entirely if it exists, with no clear messaging indicating so.

I understand the method is Read only after compiling, so I need to get used to that, that's entirely on me. Plus it would be messy anyways managing variables within Considerations.

Though, it does raise a question, since print to log statements don't seem to work within the considerations (if using custom messages from the state parameters), how can I see the scoring during run time? I feel it would help a bit to see that history. I have the debugger window open in the state tree and the blueprint debug window to watch variables, but I imagine at scale, it can get a bit harder to catch that historical information for the exact time it happened and try to piece it together.

or I am overthinking things ha

[obtuse igloo]

Anyone good at or knowledgeable with Smart Objects able to help any? 🥺

[wise sluice]

Yes except if that's about SmartObject entrance, i need to do some research in the next few days

[slow bobcat]

Not me but my other half in the AI team knows quite a lot. I can pass him the quesito tomorrow if nobody else knows the answer

[obtuse igloo]

I’d appreciate it lots!! Thank you guys 😭

[slow bobcat]

Isn't this a problem in the BT rather than the smart object itself? If I get this right:

• AI picks the item and moves to it
• player is faster picking it up
• the AI doesn't know about the player

Aren't you missing some callback for when the player picks it up?

[obtuse igloo]

I’m unsure if I’m understanding this right but the SOclaim handle is the reason why the ai moves and uses the object and how it knows where to go. I could add a delegate/event for when it gets picked up, but I don’t know I I can link it back to the behavior tree task to make the ai end the asynchronous task of it moving to the object (in the picture)

I disable the slot when it gets picked up and hoped that it would fix it but the AI continues moving to where the item was to use it even if it’s not there anymore

[slow bobcat]

You need to have a decorator aborting the move to branch when the player picks up the SO. Not sure what that decorator should check. I guess if an SO Claim handle exists for it

[heady silo]

so maybe this combined with a for loop to iterate through a list of weapon types (AR top priority, then shotgun, then pistol) before declaring a move? I don't know how to make a custom EQS query and the information I've found on the subject seems to indicate more coding knowledge 💀

or maybe a composite?

[heady silo]

This one might do for now, it basically it testing if we can see/reach the weapon most efficiently which is probably "good enough" for a first game

I think maybe a sorting method is probably doable with gameplay tags?

[open onyx]

I'm trying to call SendStateTreeEvent to trigger a state transition in 5.5. I have on-screen "print statements" surrounding the C++ SendStateTreeEvent call (so I know it's getting past that point in the code at least), and my actual state tree transition looks like the attached. The "AI.Aggroed" tag is declared / defined in C++ so there shouldn't be a typo issue. Is there anything else I'm missing? Do I need to have something for the Payload Struct value? I have a task in the Idle state (the state in which this event resides) that continually prints out "Is In Idle State", so I know at least the tree is running and the first state is active. I also have a similar printing task in the Combat state that never appears; the state remains Idle.

This is the C++ for what it's worth

DEBUG_PRINT("Target was successfully sensed.");
AggroTarget = Cast<ALegendCharacter>(Actor);
if (StateTreeAIComp && StateTreeAIComp->IsRunning())
{
    const FGameplayTag AggroTag = UGameplayTagsManager::Get().RequestGameplayTag(FName("AI.Aggroed"));
    const FStateTreeEvent Event(Tag_AI_Aggroed);
    StateTreeAIComp->SendStateTreeEvent(Event);
    DEBUG_PRINT("sent event");
}
else
{
    DEBUG_PRINT("didn't send event");
}

[slow bobcat]

Some questions:

• where and how are you declaring the tag in c++? The fact that you mention a typo rises concerns since tags are not written by hand, they are chosen from a menu. I see it selected in the picture but just in case
• where is the event transition set? Is it in the Idle state that is running?
• have you checked the state tree debugger to see if there's any extra info there?

[fierce carbon]

My zombies, when attacking, knock my character into the air. Is there a way to prevent this? I've tried to ignore impulse on damage but it's still happening

[slow bobcat]

Deleted my last message, was not for this chat or the person tagged

[elder perch]

I just want to say, thank you

[wise sluice]

Hello ! I tried to do some weird stuff with SmartObject parameters 👀 (Unreal 5.4)
It turns out that the smartObject rendering component debug display the entrance properly in editor when sliding the value but not when validating the input 😔 It fallback to its original position

It looks like it should be doable to update the debug position since the smartObject entry turns blue when it is on a right position compared to navmesh

Is there a way to have a customizable entrance location per SmartObject instance?

Okay it turns out that my actor was wrongly setup
I reparented my actor as a GenericSmartObject actor and it works well !

[chilly nebula]

Hey,

as the state tree seems to skip over the state entirely if it exists, with no clear messaging indicating so

I just tested this and it worked for me? Are you categorizing the parameter as Input ?

Regarding ways to debug the scoring, I'm asking internally.

[chilly nebula]

Pretty good overview video of the new State Tree features in 5.6:
https://www.youtube.com/watch?v=bvWHU5nvU-0

[open onyx]

1. The tag thing became a red herring, apologies. I mentioned typos to assure I wasn't getting the tag via RequestGameplayTag (and its typo-able FName arg). I have my Tags set using UE_DECLARE_GAMEPLAY_TAG_EXTERN / UE_DEFINE_GAMEPLAY_TAG in a single header/source file-pair.

I'll back up a bit and explain my setup.

• BP_NpcCharacter blueprint class. Details panel, I set the pawn's AI Controller Class variable to:
• AMyAIController C++ class / BP_MyAIController. In C++- I declare StateTreeAIComp like so

UPROPERTY(EditAnywhere, BlueprintReadWrite, Category = "AI")
TObjectPtr<UStateTreeAIComponent> StateTreeAIComp;

and in the blueprint details panel, set the StateTree asset, ST_Aggressive.

The original code I pasted is within AMyAIController's OnTargetPerceptionUpdate. This is how I have the sense configs set up:

AIPerceptionComp = CreateDefaultSubobject<UAIPerceptionComponent>(TEXT("AIPerceptionComp"));
AIPerceptionComp->SetComponentTickInterval(1);

SightConfig = CreateDefaultSubobject<UAISenseConfig_Sight>(TEXT("SightConfig"));
SightConfig->SightRadius = 2000;
SightConfig->LoseSightRadius = 2200.0f;
SightConfig->PeripheralVisionAngleDegrees = 60.0f;
SightConfig->DetectionByAffiliation.bDetectEnemies = true;
SightConfig->DetectionByAffiliation.bDetectNeutrals = true;
SightConfig->DetectionByAffiliation.bDetectFriendlies = true;

AIPerceptionComp->ConfigureSense(*SightConfig);
AIPerceptionComp->SetDominantSense(SightConfig->GetSenseImplementation());

2. The event transition is within Idle. The "debug task" is:

etc Tick(FStateTreeExecutionContext&, float) const override
{
    GEngine->AddOnScreenDebugMessage(-1, 0.5f, FColor::Yellow, TEXT("Still in Idle"));
    return EStateTreeRunStatus::Running;
}

"Running" to stop re-doing Root->Idle. "Succeeded" has same result. A duplicate task in Combat.

3. "debug0" attachment shows it entering Idle. The markers are the same, see "debug1". "states" shows the tree. Sorry for huge message 😄

[open onyx]

@slow bobcat See ^^^. Somehow it became not a "reply" after an initial error sending the message. Thx!

[simple verge]

Is there some kind of bug with StateTreeTask Run EQS Query that exits the state before other tasks are complete?

[misty wharf]

States will trigger those kinds of transitions as soon as any task completes

I think 5.6 adds the option to choose whether you want any or all to complete before anything is triggered

[lofty ember]

before 5.6? completely normal, you're supposed to handle EQS Query in a separate state and bind query result as a parameter in that state's parent state and then use it in a sibling state to the state where you handle EQS Query

[simple verge]

Is it supposed to feel inconsistent or do I not understand some subtlety? You can share a delay and a debug text task and the delay will complete, for example.

[harsh storm]

Because the debug text task probably never actually "finishes"

[simple verge]

It's also the opposite of how mass states work WRT selecting and using SO objects where the usage must be an inner state, no?

[harsh storm]

So no transition would happen

[misty wharf]

Yeah you have to design it with this in mind, 5.6 might change it I guess but perhaps it'll still be necessary to consider it for some cases

[simple verge]

Wait you're right. debug text will block completion if its the only task on a state 🤯

[misty wharf]

This is my suggestion for how to deal with this: https://zomgmoz.tv/unreal/State-Tree/Custom-StateTree-tasks#task-design

[chilly nebula]

All of this is fixed in 5.6, where we added controls for what tasks will contribute to state completion. You as the designer can toggle on/off per task, set completion to trigger on any or all tasks and so on.
Task authors can also set default, and if a task will ever contribute to completion.

FYI - Almost all tasks should complete, that is the right pattern from now on.

Regarding EQS, the best pattern is to trigger a transition on completion and read the result in a child state.

[misty wharf]

Ah nice, it also allows changing on a per task basis

I actually thought about implementing it as a bool for some tasks, whether it should finish or not, so this definitely makes sense

[chilly nebula]

one sec, I'll show you a new screenshot from today

That green + icon there shows the task is currently contributing, clicking it turns it into a - (this icon is a star in Preview release)

btw someone asked about this recently, this is what the configurable tick rate looks like 🙂 it appears as soon as you change the policy on the schema

[misty wharf]

Nice, this is a good improvement

I can delete most of the text in the custom state tree tasks page lol

[simple verge]

Some good changes for sure. I guess I was tripped up by the different behavior for FindSmartObject Mass task and the EQS Query task.

So just to be clear, the child task for Run EQS Query should not work currently and needs to be a sibling task in 5.5?

[deft kettle]

Hi, wanted to ask for some advice on how use a python RL algorithm on a agent.
My agent is a archer that can move around the room, rotate, and shoot ( these are the actions that is getting from the python script)
I am sending to the script his location, the enemy location, the distance to him, the angle between the agent and the enemy and a variable that will tell if he can shoot or not.
I am quite new in this, and I trying to create an PPO as continuous action space, basically I want my agent to learn how to shoot the enemies in the room.
My issue is that it doesn't learn anything in time, and sometimes the mu values reach Nan values, do you have any advice?

[chilly nebula]

We did some fixes in 5.6 and I have recently confirmed it works there, but I honestly cant remember which pattern worked correctly in 5.5. Did you set this up with both tasks in the same state, sibling or child state? Splitting the MoveTo to a child state is the correct pattern at least

[simple verge]

kk thanks for the help

[signal olive]

Trying to leverage the AI perception component to implement Thief-style hearing, surprisingly harder than I thought

[misty wharf]

How so? Haven't tried this but it feels like you could probably trigger sound events with varying settings based on surface or something 🤔

[signal olive]

Oh, the surface stuff is the easy part, I use an anim notify for footsteps and such as well as a data table for different surfaces for footsteps, the tricky part is the way loudness works

[misty wharf]

Ah

[signal olive]

It's still something I'm getting to grips with

[misty wharf]

const float ClampedLoudness = FMath::Max(0.f, Event.Loudness);
const FVector::FReal DistToSoundSquared = FVector::DistSquared(Event.NoiseLocation, Listener.CachedLocation);
            
// Limit by loudness modified squared range (this is the old behavior)
if (DistToSoundSquared > PropDigest.HearingRangeSq * FMath::Square(ClampedLoudness))
{
    continue;
}
// Limit by max range
else if (Event.MaxRange > 0.f && DistToSoundSquared > FMath::Square(Event.MaxRange * ClampedLoudness))
{
    continue;
}

Yeah looks like it uses loudness as part of the distance check

[signal olive]

Which I guess is okay for sounds that are actually loud enough to be heard from a distance but not so good for being able to sneak up on someone without them hearing unless the loudness is literally zero

[misty wharf]

Yeah, tbh it feels like it might be easier to just use distance. Either they hear the sound or not.

[earnest musk]

A bit abstract question - working on AI that can attack player owned structures or capture points. So far came with an idea make them visible to AI using perception, basically giving them team id. On top of that when AI sees a player structure, and sense adds it to threat list, it gets some initial threat so enemies keep attacking it if unattended. For capturing probably will go with ECS or something.
Does this approach sound sensible?

[solid spire]

Yeah that generally sounds good

[vestal briar]

Yeah, for my buildings/objects that can be attacked I have a combat component that implements IAISightTargetInterface while the outer building actors have the IGenericTeamAgentInterface setup. Just manually call RegisterSource with the AIPerception system when they come active and unregister when they get destroyed.

[slow bobcat]

Seems good, unless we are talking about many enemies and things to perceive. Then it is not. The Perception system struggles when it has to process high numbers.
How it works internally: each perceived stimuli is added to a collection (I don't recall if it's an array or a map). When processing it, the system checks how long is taking to process said collection within the frame. When reaching the time limit, it stops and waits for the next frame.
Every time said stimulis change location, new ones show up or the listener actor moves, they are added to the collection and the collection can change order.
If this happens often in high numbers, you can end up in a situation where some stimulis are not being processed ever.
This happened to me in a project. We had over 30 enemies being detected in some situations (top down arpg twin stick shooter). We had bug where some enemies were never processed (generally the ones added at the end). I got confirmation from epic (udn) that the system was never written for high volumes of stimuli happening. I ended up changing the system to disable stimulis (enemies mostly) that were further away than the vision radius.
My advice: think of a different way you can control.

• the team id is a good idea, but you don't need perception for it. There's an interface and an API for teams in unreal you can use
• structures might have different points to be attacked. Ie: is you have a castle, you probably want to attack the entrance wooden door rather than a wall. But if a wall is half-destroyed already, you probably whan to attack that. To do that, you need to check paths to specific locations in the castle. Consider options like adding interesting pints to an Octree (unreal has an octree API but it mostly supports adding, not removing. That you need to implement (it's not difficult). An octree is a super fast way to check locations within radius and obtain world data.
• every time a new player's building or capture point spawns in the world, you can add it to the octree. Your AI's can check every 2s if there's a new point of interest using the octree.
• all this can be done in eqs's that you control or in multi-thread async tasks that each AI can run making it cheap

To sumarize: if you plan on having many AI's checking on many things (other AI's, players, buildings (and within buildings, attack points), capture points... Consider an approach detached from the Perception System

If its going to be basic "yeah, the player is going to have 2/3 buildings and that's it", the Perception will be ok

[slow bobcat]

Agree with this. A good flow adding/removing or enabling/disabling things is crucial for a healthy usage of the Perception system

[harsh storm]

I keep pondering how I would handle writing ai perception 2.0 with #mass but the line tracing is where I typically stop at. Mainly due to lazinesss and not really "needing" it myself

[slow bobcat]

We don't use line traces for vision in our games. We use a coffin approach like in Siphon Filter and the lines

We do use it for hearing though

[earnest musk]

I was also thinking of switching to EQS entirely. As game base is also top down shooter with up to 30 enemies on a screen and only handful of player structures/points of defense. Need to cone up with an idea how to handle threat table for them though. But in general EQS can just run some simple tests on top of a list of players/structures snd update threat system I suppose.

[slow bobcat]

Sounds a like a better approach to me

[earnest musk]

Thanks for advices folks 🙂

[harsh storm]

I'm unfamiliar with that approach

[slow bobcat]

Game AI Pro 2, chapter 28 explains it, the vision coffin of the player
https://www.gameaipro.com/

[harsh storm]

I figured Game AI pro would've had it, just was not at my PC to look.

[harsh storm]

YOU LIED TO ME

This one I do know

I have read this one - but even so, they still use raycasts.

The sight stuff already does a distance and dot product check before doing the line trace tests. So, I don't know how valuable it actually would be to move that stuff to mass and then just tell mass to tell the actor when to do the line trace tests.

That would require the actor to update the backing entities position, which sounds like it would defeat the purpose of even using Mass.

I've lightly talked to Mieszko about at least EQS being powered by Mass as well.

[slow bobcat]

Sorry. I betrayed you!

[slow bobcat]

Yeah for walls etc yeah, but you can do some volume overlap tests etc. It's not physics free

[mighty hornet]

Hey, anyone know how to get around this error? This is being thrown when I try to destroy an AI controller running a state machine (even if logic has stopped). I assume it's because there are tasks that haven't finished being completed.

[2025.05.17-15.30.50:456][582]LogWindows: Error: 
[2025.05.17-15.30.50:456][582]LogWindows: Error: [Callstack] 0x000001a524934f5a UnrealEditor-StateTreeModule.dll!TStaticArray<FStateTreeSharedEvent,8,8>::TArrayStorageElementAligned::~TArrayStorageElementAligned() []
[2025.05.17-15.30.50:456][582]LogWindows: Error: [Callstack] 0x000001a5249f6fda UnrealEditor-StateTreeModule.dll!`eh vector destructor iterator'() []
[2025.05.17-15.30.50:456][582]LogWindows: Error: [Callstack] 0x000001a5249280f8 UnrealEditor-StateTreeModule.dll!DestructItems<FStateTreeFrameStateSelectionEvents,int>() [D:\build\++UE5\Sync\Engine\Source\Runtime\Core\Public\Templates\MemoryOps.h:102]
[2025.05.17-15.30.50:456][582]LogWindows: Error: [Callstack] 0x000001a524926855 UnrealEditor-StateTreeModule.dll!UScriptStruct::TCppStructOps<FStateTreeTransitionResult>::Destruct() [D:\build\++UE5\Sync\Engine\Source\Runtime\CoreUObject\Public\UObject\Class.h:1368]
[2025.05.17-15.30.50:456][582]LogWindows: Error: [Callstack] 0x00007ffae0c45a1f UnrealEditor-CoreUObject.dll!UScriptStruct::DestroyStruct() [D:\build\++UE5\Sync\Engine\Source\Runtime\CoreUObject\Private\UObject\Class.cpp:3736]```

[misty wharf]

eh vector destructor
It's those darn canadian vectors again

[mighty hornet]

Us Canadians always use Canadian vectors you know 🙂

[misty wharf]

I've never seen this myself and I've used ST's a bunch... have you tried looking at the callstack in your IDE?

[mighty hornet]

No, that's probably the next step though. This is happening during level teardown when I'm trying to head back to my main menu.

[misty wharf]

Are you using any custom state tree tasks or just the builtins?

[mighty hornet]

All custom tasks

[misty wharf]

Right - just makes me wonder if it's something to do with cleaning up the task which is then causing a problem somewhere

[mighty hornet]

And to make it more fun, they're all written in C++ on FStateTreeTaskCommonBase

I'm not sure if the blueprint versions have extra cleanup built in

[misty wharf]

In my experience there isn't anything special you'd need to do with them, but if you were storing something unusual on the task struct instance or perhaps in the instance data, that could do it I suppose

[wintry osprey]

Anyone knows why nav-mesh doesn't generate on geometry script made meshes? I know it works, just if there's some check box somewhere

[mighty hornet]

Nothing too unusual on the task instance data... Actor pointers, basic types. A TArray of actors and a TMap of <AActor* , float> about the most complex thing

I think I'll try debugging through VS, as you suggested, maybe set up some breakpoints around the destroy events

[misty wharf]

Yeah that's probably the best place to start, could be so many things affecting it 🤔

[harsh storm]

I still haven't really used them - maybe you have to commit them or finalize them or something?

[chilly nebula]

Are you registering to any external delegates? This looks like either a delegate firing when the ST is being destroyed, or ST trying to access an actor being destroyed (better to use weak ptrs for external refs)

[obtuse igloo]

Ive been really busy so i hadnt been able to mess with the project, but i finally got to it! The decorator wasnt checking every tick, so i decided to use a service instead! i tried adding a check into the task itself, but the problem is that the task's Receive Tick isnt executing at all. I might just have to add a blackboard key that is a bool and add a decorator anyways.

[wooden fern]

Quick question on Nav Modifiers. I have a map that is tile based, using 500x500 tiles for road tiles, grass tiles, etc.. Each tile is an actor with a Nav Modifier component to change the modifier area on for allowing where NPC's can and cant walk obviously. But im wondering if this is bad to do? Because lets say my map gets big enough and end up with 2000 tiles made up of roads and grass and sidewalks, that means there's 2000 nav mesh modifiers too. Curious on if they will cause performance issues with that sheer amount? I am using Static nav generation though. So I guess it's a one time build thing and never have to worry on it?

Pic for reference on what I am currently doing. Red is parts the AI cannot walk. The green is the sidewalks where they can walk.

All of this map is approx 1500 tile actors. (Which means 1500 nav mesh modifiers too as stated above)

[mighty hornet]

Not using any delegates, but maybe the latter? Should I be converting these to TWeakObjectPtr?

struct MRACE_API FStateTreeSetThrottleInstanceData
{
    GENERATED_BODY()

    UPROPERTY(EditAnywhere, Category = "Context")
    TObjectPtr<ARacingAIController> AIController = nullptr;

    UPROPERTY(EditAnywhere, Category = "Context")
    TObjectPtr<AVehicleBase> Actor = nullptr;
};```

[chilly nebula]

Try it and see if that fixes the issue, then update your code to check IsValid on those before using them.

[slow bobcat]

Do nav modifiers work at runtime with static nav mesh? I thought that was only for dynamic with nav modifiers only and dynamic

[wooden fern]

Yeah they seem to work. During the build process in editor it will create the static map with those modifiers involved.

But Im guessing any changes at runtime, nothing will change, is basically what I believe.

But I wonder if there is a significant cost to using A LOT of modifiers like what i've done above, using a tiled mesh map.

[slow bobcat]

Ah I see.
Sincr they are not working in runtime, aside from the extra actir per level you are streaming, I don't see any other impacts in performance

[wooden fern]

Sweet! All shall be good then. 🙂

I thought to myself, if it's static, that means its baked Nav Mesh that doesn't need recomputing. But UE will be UE, always best to ask if what I'm doing is overkill lol. It's just that my map setup is a different approach to many. No landscape involved haha

[near condor]

Hey all, just curious if anyone knows why I would be experiencing this:

If I swap the "order" of them in the State, it still triggers the Move to Location breakpoint first.

[misty wharf]

I wonder if there's a defined "intended" behavior for this to begin with

Because a lot of the mentions of tasks in states is that they "run in parallel"

So I don't know should you assume that they will run in the order you've defined them in

[rustic flax]

the fact that you can bind properties of later tasks to earlier tasks in the list implies an order, doesn’t it?

[misty wharf]

Oh that's true, I guess it would

[near condor]

Hmmm... This likely will get me the behavior I am wanting, but it feels somewhat cluttery?

[near condor]

I do feel like I frequently run into questions on State Trees, but I promise I have worked on each of my problems for a while (95% of the time)

[rustic flax]

is there a reason that StateTree doesn't allow running bits of logic attached to particular transitions? i feel in state machine code it's very typical to have like if (LostPlayerPerception) { CurrentTarget = nullptr; Cooldown = 5.0f; RequestNextState( ... ); } but StateTree doesn't permit it?

you can work-around by having a one-tick intermediate state with an Enter task, but the extra tick in between could often be problematic

[chilly nebula]

Task execution is not sequential, so you cannot have a MoveTo task right after EQS in the same state. You need to setup a child state with the MoveTo and trigger a transition once the EQS task is finished.

[cerulean otter]

So under the hood, for any single given state one should never assume any specific order for the tasks?

Was there a specific architectural reason for that behavior? What are the upsides/ downsides as opposed to guaranteeing an order (unless were flirting with threadding?)

[dim pier]

Is it possible to expose a gameplay tag picker in the state tree gui?

[obtuse igloo]

Anyone know why my Receive Abort events aren't firing for this UseSmartObject node i've got?

[rustic flax]

I think the issue here is just that you’ve assigned the variable the “Input” category, which means that manual data can’t be assigned to that variable, it can only be bound to a property

(When I first learned about magic categories and their mysterious implications I was not happy lol 😤)

[dim pier]

Oh thanks, removing the category allows me to set it now

[slow bobcat]

There's nothing in your decorator calling abort. You have implemented what to do when abort calls but not When/Why abort should call

[obtuse igloo]

I have this as the condition check, it doesnt automatically abort when this is false?

The flow control for the decorator is set to Aborts self

[slow bobcat]

Does that logic run while your sequence's children run?

I'm trying to remember something about the rules of abort. There's something related to how the nodes are setup in composite nodes I can't recall now.

[obtuse igloo]

Yes, it's constantly checking ( i didnt think it was checking on tick at first til i retried it a day later LOL)

But yeah when it goes false, its not aborting or anything idk why exactly

[slow bobcat]

yeah that's only checking the condition in the decorator. Nothing there is related to abort.
That's only a "does this decorator pass its condition?", but you are not calling abort specifically.

For example, this is what the blackboar decorator does internally

In the end, it's just calling RequestBranchDeactivation

I always work in c++ so not sure how to implement abort in BP's

[chilly nebula]

They run init and tick in correct sequence, not sure why you think they would not? I was pointing out that a task relying on the async output of another task should not be in the same state.

You can easily verify this just by printing out a string in InitState or task Tick and writing out a string set in the task.

[cerulean otter]

ah my bad, it was about async

Im still getting to grips with the whole thing and had some faulty assumptions

[obtuse igloo]

Ah I see, so far while testing, the abort event has only been called when the AI gets destroyed. I'll try and see if theres a solution for blueprint 😭 thank you though!!

[slow bobcat]

the floor you are using, does it have the flag bCanAffectNavigation set to true?

[near condor]

Thank you, this helps a ton when worded this way

[slow bobcat]

hey guys noob question about state trees: we are thinking about setting the value of a Global Param in a state tree (it will point to a behavior tree) from C++ in a component (not the state tree component).
If I have the StateTreeRef of said tree, is there a function to set the param value in it?
Haven't look very hard, but on a simple search I don't see any obvious function to do it.
I would imagine it implies something with the GetMutableGlobalParameters in UStateTreeInstanceData?

[modern owl]

For smart objects, how do you integrate the player into the mix so that if the player wants to say, sit on a bench, he follows the same check for occupancy->claim->approach->play animation flow?

[harsh storm]

It's been awhile since I looked, but I think you're right.

[slow bobcat]

Any chance you have an example at hand? And yeah... I'm in lazy mode XD

[harsh storm]

I do not. But I think EQS task does it.

I'm at my day job, so can't pop open UE unfortunately.

[misty wharf]

If your pc is not controlled by an AI controller, you would probably have to come up with some kind of custom solution for it. I don't think there's inherently anything in the SO system which would prevent it, but you would need to take it into account when building the SO behaviors (eg. you probably can't use BT based behaviors)

Creating a custom Use Smart Object gameplay task to replace the AI Task based one could make it more convenient

I guess temporarily possessing it with an AI Controller for the duration could be an option as well

[chilly nebula]

Take a look at UStateTreeFunctionLibrary::K2_SetParametersProperty here to start you off:

https://github.com/EpicGames/UnrealEngine/blob/ue5-main/Engine/Plugins/Runtime/StateTree/Source/StateTreeModule/Private/StateTreeFunctionLibrary.cpp

[modern owl]

Is having the player character controlled by an AI controller something that is done normally?

[misty wharf]

Depends on what you're doing. One case would be if you control the pc by clicking, and it automatically paths to the targeted location - this would be easy to achieve by having an AI controller possess the character.

[modern owl]

yea no not that simple im afraid typical third person

[misty wharf]

Right, if you're not using BT-based behaviors in your SO's, it shouldn't be too hard to manually claim and activate the behavior

It will require C++ though in case that's a problem

[modern owl]

nope i can do c++ its just the smart object system in particular im unfamiliar with

[misty wharf]

Take a look at the AITask for it, it's called something like AITask_UseGameplayBehaviorSmartObject

It should contain the logic to claim the SO, and to get the GB from it. Unless I completely misremember, this logic shouldn't require an AI controller

[slow bobcat]

Oooh nice. Thanks dude

[stoic pier]

I have a sight sense set up which works fine for detection and when the player moves away Stimulus.WasSuccessfullySensed() triggers as false which is what I want. If the player just stands in place and the AI walks past the detection starts but end detection never triggers

[misty wharf]

For the perception system these both should behave exactly the same, so this could be an issue with how you've set up the detection logic

[stoic pier]

found this on the forums, seems max age doesn't count down if “auto success range from last seen location” is set, changed this to -1 and works as expected

[misty wharf]

Ah, yeah I never used those myself, I've also heard the forget related settings can be kinda hard to correctly configure

[open onyx]

From my earlier post, I figured out the issue. I was putting some tasks into the "Combat" state, which had child states, but critically, none were enabled. Because of this, the Combat state is never entered even though the transition event registers (maybe a safety valve of sorts since Combat itself, has no transitions or (enabled) children, it just nopes back to root?) In any case, if the tree looks like the attached, no transition happens. If I take the Debug Text / Delay tasks and put them into an Enabled child state, the transition works.

[obtuse igloo]

Hi guys! I'm kind of in the need of some guidance on how I'd go about having an AI climb when they reach this Nav Link that I have on my wall with a custom trace channel, "Climbable".

Currently I've made it so that when the AI reaches the link, the AI's movement mode is set to fly and will trace the wall up until it gets to the top of the climbable wall and can no longer climb. I've been going off of this Gorka tutorial > https://www.youtube.com/watch?v=F_g_AnDyWi8&list=PLiSlOaRBfgkcPAhYpGps16PT_9f28amXi&index=31

I was wondering if there'd be a better way to do this somehow! Please let me know! 😊 (For reference, I'm trying to get the AI to run simple obstacle courses with swimming, climbing, and running, like in Chao Garden races)

[iron zephyr]

Hello, does anybody know how to access the currently active states in a StateTree. Ive tried some things but none of them gave me a representation of the active States. Maybe it is possible to use the Visual Debugger via cpp to get those values since it displays them, but i could not find any functions or info in the documentation.

[misty wharf]

Is there some particular purpose you need this for? There is a state tree debugger in the state tree editor which should show this, and as you said visual logger should also contain this information

[wise sluice]

If you need the currently active state in code from anywhere
You can implement this function in StateTreeComponent (or a custom child of this class)
⚠️ I coded that in 5.4, not sure it still the same in 5.5/5.6

    TStringBuilder<256> stringBuilder;
    
    if (StateTreeRef.IsValid())
    {
        const FStateTreeExecutionContext& stateTreeExecutionContext = FStateTreeExecutionContext(*GetOwner(), *StateTreeRef.GetStateTree(), InstanceData);
        for (const FName& stateName : stateTreeExecutionContext.GetActiveStateNames())
        {
            stringBuilder.Appendf(TEXT("[%s]"), *stateName.ToString());
        }
    }  

[wise sluice]

Apart from that
I'm currently playing with SmartObject + Gameplay Interaction StateTree
Anyone can confirm that in 5.5, the StateTree calling "Use SmartObject" and the SmartObject StateTree still run in parallel ?

It seems to work differently from LinkedAsset which is running the StateTree calling the "LinkedAsset state" and kind of inject the StateTree asset within it (Another point is that it is hard to debug for now)
I guess it's because it's two different systems and they were developed at different time frames.

I noticed that if i send a StateTree Event to the AI, it will receive it in the main StateTree but not in the SmartObject one 😔

[iron zephyr]

Yea, i need to display the active states in an UI Element (Maybe even a full-blown tree) at runtime.

[iron zephyr]

Thank you

[wise sluice]

If you are in 5.5, you should rely on state tags if you want to display some User Facing Text (by doing a TMap <Tag, FText>, but maybe there is a better way to achieve that)

But yeah the code that i shared should be sufficient for now

[chilly nebula]

Look at UStateTreeComponent::GetActiveStateNames()

https://github.com/EpicGames/UnrealEngine/blob/ue5-main/Engine/Plugins/Runtime/GameplayStateTree/Source/GameplayStateTreeModule/Private/Components/StateTreeComponent.cpp

[iron zephyr]

oh thanks alot

[misty wharf]

I vaguely recall it might run separate, ST's have the capability to do that much more easily than BT's (which was one of the big problems I had with BT-based SO tasks, it can't even resume the previous BT from where it left off) - easiest way to find out for sure would be to look at the GameplayInteraction Smart Object Behavior's code, it should be fairly obvious if it executes the ST there

[slow bobcat]

quick question about State Tree Event's payload: am I supposed to be able to retrieve the payload somewhere?
Let's say I send an Event with a Payload.
That Event triggers a transition to a State with a condition.
I would like to grab said payload to check something in the condition
And if the condition passes, I want to use the rest of the payload data in the Task run by the State.

Is that possible?

I'm seeing functions like

    void SetCollectExternalDataCallback(const FOnCollectStateTreeExternalData& Callback);```

and since the Payload is a ConstStructView...
```const FConstStructView Payload```

Is that how you catch the payload of an event?

I'm always super confused about how the Struct stuff works for State Trees. Any good documentation/examples anywhere to understand how is this used? It's completely Alien to me

[iron zephyr]

did you implement that directly in StateTreeComponent and built from source or separately in a child class? I just wondered whats easier

[iron zephyr]

Okay i got it working, thanks dude

[wise sluice]

Easier is a child in your project code with a child of the class
But on my side i can easily change the source code and share it so both are easy

[wise sluice]

What do you mean the struct stuff?

Event Payload struct? PropertyBinding? StateTreeTask InstanceData (which can be a struct or an UObject depending your needs 👀)?

[slow bobcat]

What I mean is: the whole State Tree system is built on top of the StructUtils: InstancedStruct (and the containers), PropertyBag, StructArrayViews and some other related code.
I have no idea how any of that works. While I can map the whole State Tree System and do complicated stuff, I fail to understand its core technology when I hit that wall.
I need some proper teaching about how the Struct Utils things work.

But my most immediate problem is to understand how to retrieve and handle passed payloads in an Event

[slow bobcat]

on this topic, the comments help but I'm still eager for a simple example (looking into StructArrayViewTest.cpp as we speak)

[summer hazel]

Hello, in the behavior tree when I use a blackboard based condition which is an enum, when I choose for example "State A" and "Not equal" does that mean the task will activate only when this enum isn't equal to state A ?

[slow bobcat]

yes

[summer hazel]

Okay thanks

[wise sluice]

Are you using 5.4 or 5.5?
In 5.4 event payload are not properly supported within StateTree while in 5.5, payload are sent as PropertyBag that you can use within sibbling/child states

[slow bobcat]

5.5. Been waiting for it to be implemented and now I'm in a good position to look into it.
AAAAh as a property bag? nice to know.
I'm debugging sending one and see what happens within the state (and its task) when called

[wise sluice]

I suppose this is a property bag like StateTree Global and state parameters
(here is a screenshot from the roadmap page : https://portal.productboard.com/epicgames/1-unreal-engine-public-roadmap/c/1494-state-tree-ux-and-events)

https://www.youtube.com/watch?v=FH2Af5s734A And it seems to be setup in Enter Condition "Required event to enter" and states should have access to the payload

[slow bobcat]

umm ok. This gives clues. I'm in c++ so I guess I would grab this info from the FStateTreeExecutionContexton EnterState within the task?

I guess somewhere here?

[wise sluice]

So your entire StateTree is in c++?

[slow bobcat]

no, but: state tree component, conditions, tasks and some other stuff is. Designers compose them using the UI like anybody else, but we handle most of the stuff in c++. It's rare we have to use anything but our custom stuff

[near condor]

Here is an example of what I have going on that is what Siggi linked to

[summer hazel]

hey guys, do people usually use the level blueprint to spawn actors when needed or create a seperate actor spawning blueprint?

[slow bobcat]

I see you are doing it from within a node (a task/condition) and that you are able to grab the property ref.
But what if I'm doing this outside a tree?
I'm in a component, I have the statetree ref and I want to change a global param in it.
How would I proceed?
We're trying something like this

if (stateTree)
{
    FInstancedPropertyBag& InstancedPropertyBag = stateTree->GetMutableParameters();
    FStructView PropertyBagView = InstancedPropertyBag.GetMutableValue();
    const UPropertyBag* PropertyBag = InstancedPropertyBag.GetPropertyBagStruct();
    if (PropertyBagView.IsValid() && PropertyBag)
    {
        if (const FPropertyBagPropertyDesc* PropertyBagDesc = PropertyBag->FindPropertyDescByName(FName("InputTree")))
        {
            if (const FProperty* TargetProperty = PropertyBag->FindPropertyByName(PropertyBagDesc->Name))
            {
                void* TargetPtr = TargetProperty->ContainerPtrToValuePtr<void>(PropertyBagView.GetMemory());
                TargetProperty->CopyCompleteValue(TargetPtr, feInstanceData.InitialActionBehaviorTree);

                stateTree->SetPropertyOverridden(PropertyBagDesc->ID, true);
            }
        }
    }
}```

But I have the feeling that we should be getting the property bag from the instance data or a on-the-go context like this
```FStateTreeInstanceData& treeInstanceData = ownerStateTreeComponent->StateTreeInstanceDataMap.FindOrAdd(stateTree->GetStateTree());
    const FStateTreeExecutionContext context(*GetOwner(), *stateTree->GetStateTree(), treeInstanceData);

[celest python]

start by learning UScriptStruct and reflection types first, and reflection in general. FInstancedStruct and its related types are reflection wrappers of UScriptStruct on surface level

oh well just saw your last message, appearently you're already ahead

[near condor]

Back before the property references were a thing, I believe I tried something similar with c++ functions, and was finding it unreliable/cumbersome to try to modify that stuff externally (Though, part of that could have been from juggling a pooling tool at the same time). If I were to try to do what you are attempting now, I feel like I would try to go about it the same way one would go about modify widgets.

So I do not over explain, are you familiar with "Widget Controllers" as a concept?

[slow bobcat]

not ahead by a long shot. Good advice that is being taken. Thanks a lot

[slow bobcat]

Widget controllers: not sure what you mean. Are you suggesting some form of AI tool?

[celest python]

Idk what ST does, but BP event graph for example is a very, very fancy runtime generated UScriptStruct that contains all function return values, local variable memory etc. - probably ST does this more programmatically but I dont think Epic "wants" you to modify things externally when it comes to ST memory

[near condor]

No, I imagine another name for them exists as the name "Widget Controller" sounds like an actual npc/player controller, but they are not. They are more like a middle man object that exists to hold variables and broadcast changes out for listeners on Widgets (so widgets don't need to hold the logic of how to get all the data it needs to display).

So what I am proposing is for an object to be spawned that is just a default, no mesh object, to serve as the "Controller", and have a global task that listens for stuff broadcast out from the "Controller", then apply the changes to the State Tree Properties/Parameters.

[slow bobcat]

yeah... this is one of those cases where I need to "pee against the wind" once again.
it should be very possible to do this if I have the Instanced data etc. I feel that we are missing a step here and we are just changing the copy of our property bag, not the actual properties in the tree instance

[slow bobcat]

aaah you mean that. Yeah that's not feasible right now. Long to explain, but not feasible.
We have a plan B, but we really would like to understand the low level of state trees as we understand the BT's low level. Otherwise we will not be able to do cool stuff with them.

[near condor]

no worries, I trust ya. Was a shot in the dark without game context

[slow bobcat]

the worst part is that I'm sure that knowledge has been discussed and solved here already, but I can't find it

[harsh storm]

Don't see how this is related to #gameplay-ai necessarily, probably a better question for #blueprint - but a different actor. I don't really use the level BP all that much.

[near condor]

idk about that, pretty heavy components are added/updated each Unreal Version as of the last few (at least in my opinion). So new stuff is added constantly. Despite that, I am still sticking with State Trees, I find them far easier to work with than behavior trees for my use-case.

[slow bobcat]

it's a case by case thing definitely.
To me understanding the flow of execution and how things are handled behind the veil is getting hard with State Trees. Check StateTreeExecutionContext::Tick and you will jump straight into the rabbit hole

[harsh storm]

https://tenor.com/o1hQ.gif

[slow bobcat]

I feel like the first time I looked at c++ looking at PropertyBag.h/cpp

https://tenor.com/bSFw3.gif

[harsh storm]

@slow bobcat Have you gotten to the capabilities talk (from Hazelight) in your GDC backlog yet?

[slow bobcat]

yeah. It's interesting, but if you are used to GAS, gives you nothing really.
Impressive gameplay iteration process though

[celest python]

well you made me curious

[harsh storm]

Funny - I heard the opposite from former Hazelight devs who work with GAS now.

[slow bobcat]

Theirs is the same but simpler. They have GA's (capabilities), modifiers (GE's), g.tags to block/allow capabilities...
They have a nice layer of Angel Script which seems to be super nice (They don't dig into it) and their "GA's" tick.
But aside that, it's pretty much a custom GAS