gameplay-ai

[celest python]

I was also participated in that conversation right? I remember something like this. But my idea was to not letting state machines affect the flow of the BT, and from what I can see your previous method was exactly obeying that rule 🤔 Main problem of BTs, you can not jump from a node to another node like state machines, and default flow logic of BT not helping radically switching behaviors. That's where you integrate different tools into BT's flow, but without breaking the BT's "prioritization" logic.

I haven't tried yet but assuming from the docs I read, your previous setup is very similar to StateTree (i.e. selectors on top of the state machines, no sequences or other composites etc.)

[visual silo]

Hmm

Yeah it's definitely feeling more difficult to abandon state-like thinking for this. The HFSM on top of the BT definitely was making it easier

Without it, I'm trying to implicitely create those same sorts of transitions through the decision logic

which I think can have its strengths but it's harder to design

[opal crest]

I can't say if you are "thinking about BT's right", since I'm still not sure if I do 🙂 But, the way I would probably start thinking about this is:

a) When will the agent "care" about sound stimuli?
b) Would the sound stimulus cause the agent to cancel any behaviour?

Then I'd put a 'HeardSoundLocation' into the BB, and add a decorator to a node in the part of the tree where I care, and make sure that the decorator can cancel any behaviour that it should override when it changes. And yes, clear it once it is no longer relevant. (it might help to have an age for the stimulus so that we can ignore sounds that are too old)

[celest python]

I have too many ideas but I'm quite lost between them

I wrote and remove multiple times 😂

[misty wharf]

Is there any reason to avoid using BT as a state machine of sorts if it works for you? 🤔

I mean it might not have all the fancy features as a more FSM-specific system but if it works it works lol

[visual silo]

Yes, this is basically what I've done. I needed to add clearing the value in the perception handling as well. It was straight forward enough to add it as a way to break out of the patrol to look at something.

Trying to make it respond to hearing something while chasing is mind bending a little right now

[visual silo]

No, not really. I do agree if it works, it works. But I'm challenging myself to think differently about it because I'm hoping it will give me better insight

and then I can go back to the state machine approach but also have a better idea of how to do more complex things

fuzzy states

[misty wharf]

Ah right, sure if it's for sake of learning why not :)

[visual silo]

The way I'm reasoning about it, it's turning into the same thinking of these nodes as lots of mini-states.

that don't really have well defined transition

[opal crest]

Yeah, that's probably a different set of behaviours using the same data, which is one of the things that BT can be pretty good at.

[celest python]

I'd not explicitly define "sound" or "sight", I'd have them as "interesting thing" or you name it. Both visuals, sounds, or other senses should feed a single abstract variable in BT and AI should be alerted in same way. If you abstract it, you can make your AI act more fluid and solve the behavior switching issues a little bit more. And I'd go more task oriented rather than having multiple BT nodes*. That's how Alien Isolation does. I wouldnt have "Attack" node either. It should be a part of your BT directly so you can prioritize it to cancel less important behaviors like investigating etc.

BTs can not jump from nodes to nodes so having abstract prioritizations help a lot to manage enter-exit thing between behaviors

[visual silo]

• So if I have a target, go to it.
• If i hear something, go investigate... and then start searching?
• If I don't have a target, and didn't hear something, go to the last spot I saw the player and then start searching....

hmm

I guess I just need to copy those nodes

There isn't a way to have two things go to the same node

[opal crest]

I think Eren is suggesting that "hearing" and last sighted could be the same behaviour. It's just a PoI that the AI should check to see if there's a delicious player hiding over there.

[visual silo]

Yeah, I could see doing it that way. There is a moment upon hearing where the agent should look and go "huh?" which is different than just continuing to chase the player

[celest python]

this could be a BT directly

[visual silo]

Although I think that's where this is breaking down too. Going to the last seen location is really just a patch for... keep going to where you saw the player until you get a new stimulus

[celest python]

But you shouldnt jump between nodes like that or copy them, i guess. I'd try to have them as more prior behaviors in BT and use a service or a BTTask to enable them by discarding current behavior

If thats not possible I'd try to copy

[visual silo]

Yep

This has been helpful 🙂

[celest python]

If you have Alien Isolation try to see its BT with this modding tool: https://github.com/MattFiler/OpenCAGE

[visual silo]

I think this points to me that I need to work more on the perception side as well to be giving blackboard better data.

[opal crest]

I have a service in my patroller that just gets the current highest priority point to check out. The service is responsible for "remembering" everything it gets from the AIPerception world and deciding which point is most interesting right now. So it's doing all of the interesting stuff like deciding if sight is more important than sound, and ranking by age, distance, and type of target seen. The BT is therefore pretty simple.

[visual silo]

Is there any reason to have it as a service in the BT rather than on the AIController's tick and perception changed events?

[opal crest]

Not really. It's just a delegation to make the AIController a bit more general and useful with different units. I don't know if that's the right decision, but it's working for me so far.

[misty wharf]

The main benefit of having those kinds of things in services is that you can use the BT to control when it's doing its thing and such

[celest python]

I'd think blackboard as memory and support/manage that memory with services, and think your BT flow as "roundtripping" logic, not states. You will switch to attack from search behavior if your AI see any player. And one way or another if your AI wont die, you'll end up switching to search or another behavior again. Try to design your blackboard with values that can support the behaviors with the info, for example if you visit the search behavior 2nd time in a very close time, and in a very close location to previous one your AI can get more aggressive etc.
That's just a mindset in my imagination of an AI, not trying to alter your AI design. Keywords are "roundtripping" "memory" and "services".

[misty wharf]

But it's mostly just a question of whether that makes sense for you or not

[visual silo]

So this BT isn't actually possible with UE's implementation?

Since there's no parallel node (that allows sequence & selectors below)

This is from that AI Arborist video

[celest python]

Its Simple Parallel

But probably not good enough compared to other engine's parallel nodes

[visual silo]

Right, to use simple parallel you'd need to combine all the logic for fire & reload into a single task

rather than have the decision logic in the BT

[ocean wren]

The way to think about it, is to consider priority at the higher level.. so what splits one priority from another. For instance, patrolling is kind of a "every now and then" thing right, so higher priority than taking a break, but lower than say "spotted something fishy to investigate"

Then think about what variables you need to break those different high level priorities apart.. the investigate thing might be "aware of something weird that needs investigating" right?

So your condition on your higher priority investigate BT would just read that value as being set to something not null

The difference here, is that if that value then gets reset to null.. you don't have to do anything, because the priority will automatically switch back to patrolling

i.e. there's no explicit state to have to transition

Same for if you spot an enemy, clearly higher priority than investigating something... etc etc..

[opal crest]

You Cheat using a service to do the parallel logic.

[ocean wren]

Want to hear something funny? I got locked out of DallE2 because I've been generating too much stuff with it 🙂

also ran out of my gdrive space saving the results 🙂

[opal crest]

haha. You did melt their 3090's.

[ocean wren]

They basically said "come back in 24 hours" 🙂

This is definitely one of those things where we NEED an open source version

but the training of that model was around 300k so unless we get some kind of distributed training going, not going to happen

[opal crest]

It seems like with ML, that models have gotten too big for hobby-ists to generate them. Yeah.

[ocean wren]

I'm waiting for microsoft to give me OpenAI access now.. so we can build some extra stuff using cognitive services.. that'll be fun!

[opal crest]

I'd totally lend my GPUs to a GTP3@home effort though.

[ocean wren]

I suspect that'll happen before too long

I mean I've got a relatively beefy GPU, but you need NVIDIA/META/OpenAI/Deepmind/Microsoft kind of grunt for these large models

at least for training

apparently this DallE-2 is estimated to run on an A100

so might just squeek into a 3090 with a bit of noodling

easily fit on my A6000, but we need it training first

[misty wharf]

There was an article a while back about these AIs and how the training cost being so prohibitive is one of the big problems

In that those who have the resources to train them can potentially gain a quite distinct advantage

[ocean wren]

It really is, and allowing people unrestricted access to them

This is perhaps the most transformative example I've seen of ML's potential for the creative sector.. its really shit that its not freely available

I've been generating storyboards for a new game as an example.. hence why I ran out of credits 🙂

[opal crest]

If training is 'embarassingly parallel' then it ought to be possible to build a global network of hobbyists to train liberated models through p2p connections. Could re-revolutionize the space if it isn't already happening.

[ocean wren]

I hope microsoft buy the rights to it and incorporate it into powerpoint 🙂

[opal crest]

Haha. That would be the one thing that would get me ever use PPT again.

[ocean wren]

Yeah, there are efforts towards that.. but getting people to coordinate is a huge ask

seriously, its the PERFECT tool for powerpoint

I suspect Adobe will do something like it.. they're big into ML now

But then they're shit at dev, so maybe not 🙂

[opal crest]

Clippy, "insert clipart of a dog dressed like a boss, designing a behaviour tree"

[ocean wren]

I mean adobe has an image hosting thing.. they could scrape the crap out of it too

[opal crest]

I can walk to their office and ask them kindly to start working on it, lol.

[ocean wren]

I was generating "aircraft carrier buried in mountain, in the style of simon stahlenhag" 🙂

[opal crest]

It's going to be wild seeing IP rights intersect with this.

[misty wharf]

and the question of how copyright applies to the training data

is the product from the AI technically a derivative work?

[ocean wren]

Yeah, the legal stuff is going to be bonkers

OpenAI claim ownership of all generated content, but I'd love to see that tested in court

especially when you get multiple models up to the same quality (imagen from google recently for instance)

One of mine from yesterday

And another.. for this storyboard

[misty wharf]

nice

my friend was using dall-e mini to generate some parts for some promo images he's making for some new music he's releasing in a couple weeks

[ocean wren]

Trying some different things

I quite liked the sort of steel round thingies..

One thing that pisses me off, is that it generates these amazing images, but often crops off the heads, or crops part of the round image 😉

Like this one

I haven't figure out a way to guide the prompt to zoom out a bit 🙂

I guess thats why we still have to research what prompts actually do

But the point remains, I couldn't in like a trillion years create these images myself

Well, not in like a few seconds 🙂

[opal crest]

That's some game concept art if I've ever seen it.

[ocean wren]

Its really compelling being able to imagine something, describe it and see it

I was considering posting on indiedb with a bunch of this stuff just to see the reaction 🙂

[opal crest]

Would be incredible if they could return a parameter space explorer, so you could see what was "nearby" but you didn't quite bullseye with your prompt.

[ocean wren]

There are a few things you can do in the interface

like you can mask off parts of the image and give it a prompt to generate the masked off areas

So for example:

That's Jake Birkett, indie guy, posts images like this every now and then on twitter

so I asked his permission to try this...

So mask off parts and add "post apocalyptic" to the prompt

[opal crest]

Oh, that's crazy.

[ocean wren]

Or change the prompt

[opal crest]

Would that work with any source image? Kind of a hybrid prompt?

[ocean wren]

Yeah, I found it a bit.. picky.. it didn't do well with like full head shots

I think it needs sort of enough space to generate the infil

There was a good example on youtube.. where someone cropped out a bunch of plants, then generated some ghibli trees, then masked only the trees and generated the terrain 🙂

[opal crest]

Right. It's a tempermental artist, needs enough canvas to express itself.

[ocean wren]

Exploring the prompts and how to guide them will be a new creative skillset, mark my words

you CAN ask it for variations though.. like this:

So you can find something you like, ask it to add variations

[opal crest]

The secret language thread was really interesting. (that gibberish had a semi-consistent meaning),

[ocean wren]

Yeah, a big part of this stuff is figuring out how prompt weightings work, some keywords nuke my generations 🙂

robot basically kills anything good 🙂 cos they pull in a ton of shit looking stock art robots 🙂

[opal crest]

Can you prompt the variation? "This, but sadder"

[misty wharf]

I'd try some other words for robot-like things, like mech or cyborg

[ocean wren]

No, its not quite that directable, you can redo the same prompt and add sad, but then sad might change the overall look a lot

yeah, I did try mecha.. lemmie find it

[misty wharf]

I wonder how does it work with non english languages

[opal crest]

Or "gundam" if you want to force a certain style, I imagine.

[misty wharf]

if you used メカ　for example which is mecha in japanese

[ocean wren]

Yeah, that might actually work a lot better

I used mecha

One problem I've got, is that I broke out each image into a directory with the prompt used to generate it.. but i've done thousands of images 🙂

[misty wharf]

:D

[ocean wren]

stuff like that

Those were both mecha prompts

[misty wharf]

ロボット or ロボ might be worth a try as well (robot and robo respectively)

[ocean wren]

I'll give that a go tomorrow night when I'm allowed to generate again 🙂

Here's something funny..

Lookit the hands 🙂

For some reason, it really doesn't like female ninja hands 🙂

This one here though...

I mean that one is fookin nice

It really does generate some nice poses

[opal crest]

I'm happy that algorithms also hate drawing hands. There is balance in the universe.

[ocean wren]

I guess training on hundreds of millions of concept art pieces counts for something 🙂 apparently its dataset was 460million image/text pairs

I mean its not generating anything NEW, its all derivative.. but I suspect thats like 80% of commercial art, it not more

[opal crest]

So, there's definitely some lab out there trying to tune parameters for doing this with 3d models, right?

[ocean wren]

And frankly, I've seen people trying to charge money for worse 🙂

yeah, I suspect that Epic will do it with Megascans once they reach a bigger database

But there's already a dataset called shapenet that we use for generative models on 3D objects

Did you guys see InstantNerf btw? 🙂 just a heads up.. thats the future of rendering too 🙂

I guess I do like the whole dystopian thing way too much 🙂

[opal crest]

It also pairs well with generative art. Discontinuities or a sense of the broken is thematic.

[ocean wren]

Yeah, exactly

some of the more.. atmospheric images have that nice sort of "not photographic" thing

[opal crest]

Instant Nerf is way better than I thought it would be. Image based 2d to 3d has come a long way.

[lyric flint]

Can anyone help me understand player controller and player indexes?

[ocean wren]

so it gets the basics, but representational rather than an absolute rendering

[lyric flint]

I don't see the correlation

[ocean wren]

there isnt any

player index is just the index of the player in the connected players list

player controller is a class that a player uses to control a controllable class (character usually)

[opal crest]

Ah, this is two years old, and looks to be the equivalent of "this person does not exist" https://github.com/marian42/shapegan for the shapenet dataset.

[lyric flint]

so do you need a player controller for every player?

[ocean wren]

yeah

and an AIController for every AI

[lyric flint]

but how does the player controller know which player it is "connected" to?

not the pawn, but the actual player

[ocean wren]

Yeah, if you're interested in sort of 3D graphics ML state of the art, look at Michael Neissner's youtube channel: https://www.youtube.com/c/MatthiasNiessner

Michael basically has all the top profs in ML 3D generation on his channel, basically records seminars for his students

I quite like some of the work by Andreas Geiger

and his students obviously

Neural heads 🙂 https://www.youtube.com/watch?v=S7LY1DtJCsI

[opal crest]

This is great. I have some fun catch-up to do here.

This is getting closer to the sort of things I'm curious about: https://www.youtube.com/watch?v=ZqgiTLcNcks

[ocean wren]

I've got a soft spot for Michael Black's lab and his work.. very similar to my own interests

[opal crest]

AI mediated kitbashing would be a great gamejam/indie prototyping tool.

[ocean wren]

https://www.youtube.com/watch?v=YVxRNHmd5SA

There's a load of work thats sort of on the cusp of usability

things like depth from rgb images.. instance segmentation and voxel creation

Thing is, a lot of this work is good for robotics guys too, so there's a lot of interest in increasing sota

[opal crest]

There seems to be a lot of focus on RGB -> 3d.

[ocean wren]

So I'm just on the lookout for stuff we can use in games 🙂

[opal crest]

Which, yeah, robotics and self-driving.

But also Camera -> In game model.

[ocean wren]

I'm toying with a prototype over summer, where we generate an Unreal Engine scene from a camera image (i.e. we layout the scene given an input jpeg)

[opal crest]

Though I pity whoever is trying to convince the AI to generate bones.

[ocean wren]

pose estimation is probably the easiest part now 🙂

[opal crest]

Right, for people that assumes a skeleton, and maps input to bone parameters based on image processing?

[ocean wren]

it basically predicts a probability function for each pixel of a given joint

classic lib is CMU's OpenPose, but its a bit crap

[opal crest]

I was thinking more of given enough footage of real thing, determine its points of articulation and construct a skeleton.

[ocean wren]

Oh there are those too, look at Michael Blacks lab for that kind of thing

look at the smpl-x work

https://www.youtube.com/watch?v=rzpiSYTrRU0

[opal crest]

Man, I remember when I was excited about just estimating Gaze from a camera when the person was wearing IR markers.

That was a fun grad project though.

[ocean wren]

Yeah, this stuff is basically easy now.. Google's MediaPipe does a pretty good version and is fun to play with

https://www.youtube.com/watch?v=K-_krrs05qw

Its not quite mocap quality, but if you multicam some of it it improves

Epic gave a megagrant to a company doing exactly that about a year back

Move.ai I think they were called

[opal crest]

Yes, I have a fonder on my hard drive when I was taking some notes on whether the free tools were good enough to do some simple home mo-cap.

Answer was 'not quite' and I wasn't excited enough to revisit my signal processing notes to clean up the input.

[ocean wren]

https://paperswithcode.com/sota <-- usually has the state of the art with code 🙂

You want monocular 3D pose estimation

Facebook is doing a LOT of this kind of thing: https://www.youtube.com/watch?v=eBf_SzIgIVI

[opal crest]

There definitely look better than what I was able to dig up last year. Still an essential tremor that is pretty common to camera based methods, but there must be some post processing that I could do.

[ocean wren]

Yeah, for sure, you could also add a stability term into the model itself

I've seen RL versions doing that to good effect

also adding in physical properties like anti-self-penetration helps etc.

The jittering is probably the easiest thing to deal with for slower motions at least

[opal crest]

Yeah, the silver hammer I used to use for things like this was a kalman filter, but that's going to make all motions smoother than they probably should be.

[ocean wren]

yeah, there was a filter method that motionbuilder used to use.. but the name isn't coming to me.. something like tubbman or something

I'd have to search the literature

I'd go for a kallman initially though

[opal crest]

You've been incredibly good at indulging my curiosity 🙂 Thank you.

[ocean wren]

maybe scale the kallman based on joint velocity or something

No problem man, I like this stuff myself 🙂

[atomic coral]

omfg

[ocean wren]

We're installing a mocap studio at work soon.. so will get to play with a bit more expensive kit, but I want a monocular thing at home

[opal crest]

Yeah, the prediction parameter is probably going to be "whatever the current rotational velocity is, continue that" so it'll smooth out changes to that.

[atomic coral]

that whole time i think i was instancing the FMyRecastQueryFilter wrong.

[ocean wren]

Oh, that won't help 🙂

[atomic coral]

managed to get the query filter actually hitting the proper filter functions, by changing InitializeFilter

[ocean wren]

Good work 🙂

[opal crest]

👏 🎉

Did it require a source build in the end? I never did finish the thing I had that was supposed to annotate paths with the modifiers it passed through, and if your work works without a source build, I might have to dig it up.

[atomic coral]

i'm still trimming it down to a minimal working case, but it looks like to make a truly working custom nav filter, you just need to:

• subclass FRecastQueryFilter
• set bIsVirtual = true in constructor
• override passVirtualFilter and/or getVirtualCost
• subclass UNavigationQueryFilter
• set bIsMetaFilter = false in constructor
• override InitializeFilter to:

Super::InitializeFilter(NavData, Querier, Filter);

[opal crest]

Awesome! That's in my notes now. Incredible work.

[atomic coral]

now to figure out a way to pass info about the pathing entity

now that i've got a minimal case, i posted a quick little summary onto the forum thread that inspired me
https://forums.unrealengine.com/t/modify-pathfinding-cost-detournavmeshquery-subclass/342472/13

[atomic coral]

then the big question is.. how to actually make use of this data

[mint terrace]

that is awesome that it's working

can you get to the querier with what you have or

actually looking at the polys and understanding the prev/current/next poly relationships will be the meat of it, but before that you'll need the querier I imagine

[atomic coral]

I think it should be sufficient to just store a TWeakObjectPtr to the given Querier at initialization . . .

[mint terrace]

yeah seems fine?

maybe I'll do one of these, I have ideas about some runtime cost stuff I'd like to do

[atomic coral]

i'm hoping it'll be sufficient to compare getVirtualCost's pb parameter to the location/rotation of querier, but i'm just getting to checking out the data in pa/pb right now

[atomic coral]

hmm. how to cast the Filter.GetImplenetation() up to the actual class, so I can call a function on it......

Casts is erroring i can't convert INavigationQueryFilterInterface to FMyRecastQueryFilter ..

FRecastQueryFilter implements INavigationQueryFilterInterface
FMyRecastQueryFilter extends FRecastQueryFilter
Filter.GetImplementation() returns *INavigationQueryFilterInterface
Cast<FMyRecastQueryFilter>(Filter.GetImplementation()) errors 'TCastImpl<From,To,ECastType::UObjectToUObject>::DoCast': none of the 2 overloads could convert all the argument types 1> with 1> [ 1> From=INavigationQueryFilterInterface, 1> To=FMyRecastQueryFilter 1> ]

[mint terrace]

ahmmm is there special casting for this sort of interface cast

yeah I think it's GetObject() and Cast<>() that?

no wait that's TScriptInterface

[atomic coral]

i shouldn't have to declare that my subclass also implements that interface.. but i wonder

[mint terrace]

/shrug

[atomic coral]

nope that errors that it's already a base-class

derp, static_cast is probably the answer, Cast only works on UObjects

Things I've never needed to use before for $100

[mint terrace]

https://giphy.com/gifs/jeopardy-the-jeopardy-you-know-2yvoIFyZghBDszbIk3

[atomic coral]

lol in the thumbnail version, i thought the category was "mierdes" .. i may need to see an optometrist. But "mierdes" is definitely a good category.

QuerierLocation=(X=10840.777 Y=-10312.226 Z=99.159) 
pa=(X=-10840.777 Y=10312.226 Z=10.000) 
pb=(X=-10845.000 Y=10080.000 Z=10.000)

[mint terrace]

off to the races

interesting that X is inverted?

maybe that makes sense though in the level i dunno

oh wait Y is too

[atomic coral]

thanks to some poor souls previous notes in unworking code that Y and Z are reversed in Detour...

[mint terrace]

lovely

oh i guess pa/pb are relative to querier location

maybe

[atomic coral]

according to the doc for Detour (holy wow, real documentation!) pa and pb are the start position on the edge of the previous/current poly, and the end position on the edge of the current/next poly

so, yeah, it does look like to make sense they'll need to be inverted

[mint terrace]

which docs?

need some leisure reading materials

oh yeah its just the regular stuff https://docs.unrealengine.com/4.26/en-US/API/Runtime/Navmesh/Detour/dtQueryFilter/getCost/

i just figured it would be documented on getVirtualCost but

since thats just a way to override getCost() anyway, this makes sense

oh tbf there is a See Also on getVirtualCost

[atomic coral]

((GetActorLocation() - MoveEnd) | GetActorRotation().Vector()) should give me > 0 if ActorRotation is facing MoveEnd, right

[misty wharf]

Probably

You might need to normalize the direction tho, not sure

[opal crest]

Dot product should work properly without normalization.

[atomic coral]

i guess GetForwardVector would be better

[atomic coral]

i suspect i'm doing that wrong. i need to do these calculations more than once every 5-6 years to remember how they go lol

[atomic coral]

hmm. it's not having the overall effect that i was hoping for. setting it to go to a target that is 500 units behind him, with three possible exits, one behind, one in front, and one to the side, i expect him to not pick the one behind every time

[opal crest]

Are you only using the dot product? Or are you combining it with the original cost function?

[atomic coral]

multiplying the original cost times 1000 .. but then it applies some hueristic thing since it's a nice big poly that it's on, and reduces that in half or so. i'd still expect that to cost WAY more than taking the shorter path that i'm looking at, though

[opal crest]

I see, so when a path segment heads behind the agent, you multiply the cost by 1k? And it still takes that path?

[atomic coral]

yes. i wonder if there's some way to visualize pathing costs

[opal crest]

You could use debug draw?

It could be a mess, but drawing each path segment considered and the cost associated would be readable. (since it should compute the same cost each time, even if it visits the same segment more than once)

[atomic coral]

i'm going to try returning DT_UNWALKABLE_POLY_COST and see what happens.. that should completely eliminate that path from consideration....

lol.. ok.. now he turns, walks to it, but won't take it, and now the other exits are invalid

... i should probably check that the other exits even work

[mint terrace]

i feel like gameplay debugger should have this (visualization of costs)

but for all its power, it kind of misses the boat on that stuff

also the gameplay debugger lies-- if you focus a character, then change its controller (Possess), it will report no controller

fun fact

[atomic coral]

gotta be something i'm missing out on here - now he's walking to the forward exit, turning and going to the one he shouldn't use, in the same move

it's literally pathing him to the far end of the poly then back to the near end

i'm starting to suspect my idea is rife with holes that i didn't understand when i conceptualized it, and still don't

[opal crest]

You might need to store the initial facing direction instead of the actor.

The navigation system will recalculate from time to time while you are moving.

I'd be curious if it works well at drawing a path when you just call FindPathToLocationSyncronously

[mint terrace]

i assume there's going to be lots of cost calculations, from one poly to the next

[opal crest]

since that'll only fire off once

[mint terrace]

so you cant just affect them all

at least, the same way

youd need to factor in where in the path it is and how you might "plan" that route from where you start to where you want, and then yeah as epigraph is saying, expect that it will recalculate that as the agent navigates the original path

[atomic coral]

yea i'm checking to see if it's doing any re-querying

[mint terrace]

if you executed a single moveto, and just observed what calls getVirtualCost() is getting, i would assume you get 1 call per nav poly boundary

so is current where the agent* currently is right now, and prev/next signify the boundary that is being checked? aka, current has no direct relationship to prev/next-- so for all the calls to getVirtualCost() in the first round of navigation planning, youd get current being the same poly/tile/etc and prev/next would be iterating over the nav poly boundaries being considered?

[opal crest]

I've been thinking. If this is really about exits and you only have a player agent, you could do something like "when player enters the room, mark the exit they entered from with a navmodifer_obstacle. (you could use a box collider the same size of the navmodifier to track the last entry the agent took)

[atomic coral]

ok, conceptually, it works, if the front exit is pretty close, he'll take that instead of the rear exit, which is the much closer route without accounting for turning

[mint terrace]

i think the exits are just a way to constrain it to check behavior (during testing)

[atomic coral]

i could make it so that the smart links for entrances and exits close .. but this seems like a much better way once i figure out the thorns

[opal crest]

It is. It's a very cool approach.

[atomic coral]

if it's recalculating on the fly, it's not going thru my filter

[mint terrace]

i think it will only recalculate if needed

which would be if the state of the navmesh changes (like modifiers change etc)

would also include smart links changing

etc

[atomic coral]

hmm.. i don't see how any of the other data is especially useful.. right now i'm just using pb, which is the "end position on the edge of the current and next polygon"

[atomic coral]

hmm.
ok, so. MoveEnd is 'pb'.
Dot is the Dot Product of Rotation and Location-MoveEnd
If MoveEnd distance from current location < 200, then if Dot > 0.8, return maximum cost. If Dot > 0, return regular cost * 10k, otherwise return regular cost.

[ocean wren]

There is a backtracking thing in the nav system btw.. saw it in the docs yesterday, no idea what its about though.. just thought I'd mention I saw it

[bleak valve]

Hi are there any tutorials for neural networking in unreal engine

[short smelt]

Hello everyone, is there a bug on the Move To node in the Bahavior Tree on UE5? I cannot plug actor keys into, only the SelfActor shows up. (Also transform keys)

[misty wharf]

Make sure your object keys have their class set to actor

It will not work with transforms, use vectors instead

[short smelt]

Like this right?

[misty wharf]

Click the arrow next to key type to expand it

This will let you specify the type of object

[misty wharf]

the other arrow

[short smelt]

Ohh

Thanks!!

[ocean wren]

There are not.

[ember sandal]

hey I am brand new to the UE AI world, so I hope my ignorance isn't too baffling lol. But I have a (hopefully) simple question. Is there a way to do an AI assisted MoveTo without a NavMesh?

[misty wharf]

no

you can certainly implement something else yourself but I'm not entirely sure how you intend it to work without navigation

[ember sandal]

my goal is to have ships fly around in space

but flying doesn't seem to be what the NavMesh is built for

unless I just don't get it

[misty wharf]

Yeah it isn't really

[ember sandal]

so there is no way to get AI assistance in flying?

[misty wharf]

Basic flight in space is quite simple though, just move the actor towards the point over time

There is no flying pathfinding out of the box no

[ember sandal]

I see what you mean

so I'm overthinking things with this

trying to over engineer

[misty wharf]

Yeah it really depends on the specifics. If you don't have obstacles you need to avoid it's quite straightforward

[ember sandal]

yea I don't really have any obstacles to avoid, its just a matter of going from one place to another

@misty wharf thanks a lot for the help

[misty wharf]

👍

[fossil wraith]

should i be doing AI in CPP or Blueprints I'm still procrastinating and I just don't know which one I should use as my AI is planning on being complex. ive asked this before but i also just dont know what to do or how to do stuff. Im making a stealth game AI.

Ai functions i need:

Path system

Ai to follow the path

Each Path point needs to have options such as whether it is just a wait type or it does an action and if i choose action it needs to give a list of what actions i can do. and obviously with the wait function i need to put in the amount of time.

Obviously Basic form of detection which i can use the perception system and see if the AI spots an actor that derives from my player class.

And i also want the AI to stop following the path after 1.5 seconds of seeing the player then the ai will either chase the player or go to the players last known pos.

I also want a way to get the AI to notice different things like dead AI (havent implemented yet as i dont know how to do it yet)

[misty wharf]

If you use BT's it's not hard to replace them with C++ based tasks later if it seems useful

even if you aren't it really doesn't make that much of a difference and is possible to change later

[fossil wraith]

i cannot find any good tutorial or resource on what i want my AI to do

[fossil wraith]

do you know any decent resources that will cover what i want to do?

[misty wharf]

There's been some discussions on this channel previously about stealth game mechanics and I think there might be some info on youtube but you'll just have to search for it because I don't really know them off the top of my head

Most of what you want doesn't sound too complex

[fossil wraith]

i think most of them was me tbh but im just so confused and im getting really really stuck

[misty wharf]

I would suggest trying to break down your problems to small parts and then going step by step

For example the AI to follow a path and have actions at waypoints, you can just start by making your AI move to a specific point and then make it follow a series of points, and so on

[atomic coral]

why can't you move without navmesh? uncheck "use pathfinding", and if you're in a mode capable of movement, it should work

[fossil wraith]

why would you want to move without navmesh?

[atomic coral]

if you're using character movement, or something derived from that, anyway

if a walking character finds itself off navmesh, or suddenly flying or falling

[atomic coral]

probably "yes". I find Blueprint to be crazy useful for Behavior Tree implementations, even though I generally prefer to write in code where I can. I'm a big fan of Behavior Tree, now that I have a much better understanding of how it works

[atomic coral]

i don't know if the current RecastNavmesh implementation supports it, but in a prior version of Unreal, I know of someone who set the navmesh height quite high, and basically treated it like a volume instead of a plane

[bleak valve]

Oh but do u know how to do it

[opal crest]

I followed in your footsteps and built a visualiser for this approach. It's interesting how rarely the extra weight for direction seems to matter unless you do what you did and mark it as not traversable. If you are going around obstacles you tend to both go with and against the grain, so it balances out. Haven't figured out how to get the prior path segment so I can try to make it about changes in direction and prefer a straighter, if longer path.

[ocean wren]

Yes, but you won't like it 🙂

[bleak valve]

Haha but could u give an idea about it

[ocean wren]

It depends 🙂 what is it you're interested in doing?

[bleak valve]

I need a neural network for throttle control for my vehicle

[bleak valve]

This ^

[ocean wren]

Riiiight.. in that case my answer is overkillorama for you 🙂

You basically need like a single layer perceptron 🙂

But just for the record, my recommendation is to make your NN's external to UE and call them from a HTTP request

[bleak valve]

Oh

[ocean wren]

For your specific use case, I'd just hand code a simple NN in C++ and have done with it

[bleak valve]

Should I use raycasts to detect collisions

[ocean wren]

There are people who'd done plugins for PyTorch and the like, but honestly its too much pain 🙂

[opal crest]

If you had suggested embedding python. I may have rioted 🙂

[bleak valve]

Isn’t PyTorch only python?

[ocean wren]

That's a bit of a difficult question.. again it depends on your game

Pytorch is C++ too

typically you'd create your model architecture in python, then save it and then load the saved model arch in the C++ version

[bleak valve]

So ur recommending PyTorch for this use case right

[ocean wren]

But again, that's totally overkill for a throttle control

[fossil wraith]

wait you can use python with UE?

[ocean wren]

No, my recommendation would be to quickly do a simple NN in C++ yourself

[ocean wren]

Yes and no 🙂

[bleak valve]

Ok seems complicated lol

[ocean wren]

You can use python in the editor

But for some mad reason, they didn't enable it for the runtime

[fossil wraith]

so python for editor stuff but not gameplay

[ocean wren]

Which is strange, because another group made it work at runtime too, but gave it up when Epic released their editor only python

its weird

[bleak valve]

As u said there’s a plug-in for PyTorch in UE right?

[ocean wren]

Epic added python I suspect because its required in the VFX industry

Yeah, look on github

[fossil wraith]

blender uses python i know that and I think Maya and Max can as well

[bleak valve]

But I am not sure how do I start

[ocean wren]

Yeah, python is all over the place, hateful language 🙂

Pro: can't help you there.. plenty of tutorials out there on the interwebs about pytorch

[opal crest]

Perceptron's are not complicated. The terminology makes them seem that way, but it's really worth coding one up outside of unreal to see how NN's work.

[bleak valve]

Thanks for your advice

[ocean wren]

what he said.. agreed.. make your own.. its pretty simple

you don't need GPT3 for a throttle 🙂

[bleak valve]

I can use tensorflow also right?

[ocean wren]

Yeah, pretty sure TF has a C++ build, but it uses their crappy bloody build pipeline, so probably not UE friendly unless some pyscho has taken the time to figure it out 🙂

[bleak valve]

Ok I’ll start with a basic neural network outside ue

[opal crest]

https://github.com/getnamo/TensorFlow-Unreal There is this, but I just googled one up just now.

[ocean wren]

Again, you can use HTTP requests to an external model.. I'd write a python thing in PyTorch, host it via fastapi and then hit the fastapi with calls via HTTP requests in UE

[bleak valve]

Last question: do games usually use neural networks

[ocean wren]

The reason I advocate that, is that you can test it more easily and honestly its probably a wiser move to run the NN on an external GPU

That's a bit hard to explain, but in general, no.

But yes, some do

Or rather, they do for certain tasks, not always for gameplay

A load of games-as-a-service use ML for things like customer prediction, attrition, advertising, fraud detection and the like

Some have used them for training bots

[bleak valve]

Oh

[ocean wren]

I use them for cinematography

if you search for "GDC machine learning" on youtube, you'll see a bunch of examples

Lots of people doing ML for animation stuff right now

[bleak valve]

Thanks for your suggestion

[ocean wren]

actually. Epic just added some ML based animation deformation stuff to UE 5.0.1

I wouldn't recommend it as a beginner thing though 🙂

The Embark guys used reinforcement learning for balance controller on Ark Raiders

[bleak valve]

Well then I guess I shouldn’t use it

[ocean wren]

so yeah, its kind of a bit niche still

[bleak valve]

It’s gonna be my first neural network

[ocean wren]

In that case, I'd recommend writing your own

in C++

learn how simple it really is

[bleak valve]

There’s a lot of math involved

Apparently

[ocean wren]

pfft, its only simple

dot products and the like

numbers go up, numbers go down kind of stuff 🙂

[atomic coral]

if you do happen to come up with any interesting results, i'd like to hear. getCost receives a prevTile and prevPoly, but I don't see how any of the data in the dtPoly and dtPolyRef structs are even kind of useful (this may be my lack of experience)

if i don't come up with an acceptable result here pretty quickly, i'm going to have to shelve this work, admit temporary defeat, and go work on other tasks for a while, and probably tackle the problem from a level design standpoint rather than code

[ocean wren]

Well, you can get neighbours from the dtPolyRef, which might be useful

[opal crest]

It might, though I'd really just like to be able to get the current candidate path (if that's even how the alg works). The filter might not have enough info about the state of the search to do that.

[atomic coral]

@ocean wren i'm not seeing anything that looks useful -- perhaps some of the indices/handles to other data might be able to provide further information, but i don't know where to plug them into?

[opal crest]

The we could add a fixed, but non-linearly increasing cost to turns, and let the g term of the A* search nudge the path toward one that takes gradual turns.

[atomic coral]

@opal crest nope, i think the most state information you're going to get there is "here's the poly i'm looking at, along with the previous and next to be considered for this path" .. at least, i'm assuming it's the previous and next for this calculation

but without any position information for the prev/next, i don't see how it's useful.. unless one can dig other structures from that data

[ocean wren]

You really need to do a dyjkstra rather than an A* don't you? so that you can determine the cost of routes

Just feels to me that trying to do it from heuristic cost is missing something

[atomic coral]

That may be right -- I've honestly never researched pathfinding, what's been given to me in Unreal has always been sufficient before, with a little nudging here and there. Definitely implementing a whole thing like that would be out of scope. The time I've spent on this already is out of scope, but I thought it would be worth it for future exploration also

[ocean wren]

Well, its good to learn this stuff 🙂

[opal crest]

My intuition is that there isn't any difference between adding g (path cost) to discourage a route because it's difficult terrain, and adding g to discourage a route because it has turns in it.

h definitely needs to remain untouched (since it measures how close you are getting to the goal).

[ocean wren]

You might be right.. but my pathfinding spider sense is tingling

[opal crest]

Yeah, and this is definitely weird territory.

[atomic coral]

well if nothing else, i've sent some people down a .. erm... path.. of discovery 🙂

[opal crest]

One thing I observed is that because the total angle to get to a place, if you don't have a maze, is generally the same between two routes. So just adding up the angles and assigning a cost doesn't change the result.

Which is why increasing it non-linearly seemed like a solution, since it would penalize sharp turns more than gentle ones.

But then the topography of the navmesh would start to matter quite a bit.

It might just be that we want to only penalize the angle of the first segment, so that the character doesn't head straight back, and then otherwise score it normally.

But yeah, it's been fun revisiting classical pathfinding because of this Q 🙂

[atomic coral]

honestly, since we're traversing a link and transferring control to a different BT Task, immediately after hitting the end, we're doing another pathfind anyway, so I don't expect long paths to matter too much

hmm. not sure if this is useful, but once the exit the ai is facing becomes farther away than the one it's not facing, then it will start backtracking to the bad one

the mindboggling part is it walks all the way to the good one, then reverses

[ember sandal]

so I tried this approach yesterday actually, and the pawn didn't move. that's when I came here. I had a feeling I may have been missing some other configuration and thats why the movement didn't work tho.

[atomic coral]

I'm pretty sure it should work -- although i wonder if you maybe need to HAVE a mesh for it to work. Are you using CharacterMovementComponent?

[ember sandal]

No I am not. and thats why I'm sure I'm doing things wrong. At this point I just have a blueprint class derived from the Pawn class. And there I have a custom Playercontroller attached

[atomic coral]

ok, my experience is using AI controlled pawns that are derived from Character, and use CharacterMovementComponent

[ember sandal]

ahh I see

I think thats where my gap in knowledge is. I don't have any "characters" in my current game just ships and vehicles

I am now feeling like I have gone about everything wrong

[misty wharf]

MoveDirectlyTowards might work for floating pawn movement or such

There's no reason for it to be a character if it isn't a character

imo

[atomic coral]

ok i'm not familiar with what is and isn't in the MovementComponent, but I do know that there's a lot of psuedo-physics handling over in CharacterMovementComponent

[misty wharf]

Yeah, if it's for moving spaceships it seems like kinda pointless :)

[atomic coral]

right, but something like the Flying movement mode might be useful to look at. I'd wonder if the AI movement stuff has a soft dependency on CharacterMovement.. like.. not a real dependency, but it doesn't actually work without it

[ember sandal]

hmm I see

[misty wharf]

iirc it doesn't specifically require CMC, it uses PathFollowingComponent

But I think it requires some variety of Movement Component

[opal crest]

You can derive from Pawn Movement Component, but you will be doing a lot of the work yourself.

[atomic coral]

that does sound correct, but also the MovementComponent might have no idea how to move the Pawn to comply with the PathFollowing

[misty wharf]

Yeah

FloatingPawnMovementComponent might be able to, ProjectileMovementComponent might not :)

[atomic coral]

i'm.. also talking about things i only have a cursory understanding of, as all of my recent work has involved already existing characters, so my knowledge is incidental to my experience

[ember sandal]

this has been an incredibly enlightening discussion for me

I wasn't sure which direction to move in, but I think I know what to do now

[opal crest]

FloatingPawnMovementComponent is a good start for spaceships, though you'll need to add logic for turning, etc...

[ember sandal]

sounds good because thats still a better place to start from than where I am now

[atomic coral]

oh, there actually is a FloatingPawnMovementComponent 😮

[opal crest]

I don't remember if it responds to AIMoveTo properly though (even if your pawn has an AIController)

[atomic coral]

I really miss having an editor that can easily do "Show me all child classes of". Boggles my mind that that is lacking from everything

hmm. even setting that cost as unwalkable... it still applies the heuristic, which reduces the cost, which makes it not ignored..

[opal crest]

Yeah, ultimately it's going to be balancing 'distance to goal' with path cost, and taking the greediest option. Weird that it overrides unwalkable though.

[atomic coral]

it's in the 'special case' in the function that calls getCost, for the last move in the path.. since it's also the only move in the path

i think

my getCost is only being called twice in this test level, one for each link out of the poly the AI is in

at least until the AI gets to a larger navmesh area

[opal crest]

I see, so there's really only one nav poly.

[atomic coral]

effectively, yes. it's standing on a poly, that has a pair of navlinks out of it

[opal crest]

And it's just trying to find "out", and somehow that ends up just being 'closest' because of the special case for last move.

[atomic coral]

The 'pb' that i'm getting for each call is the exact location of each of the navlinks entry points

[opal crest]

That makes sense. It's normally drawing a path from the entry point of the poly, to the center of the next poly and pb is the point on the edge of the poly that intersects, and with the navlinks case, I guess that's the navlink location.

[atomic coral]

i take that back, it applies the heuristic, but then hits the continue anyway, because it's comparing the unwalkable score to the variable pre heuristic

so.. why the hell is it going there?

[ocean wren]

It intersects each neighbour edge and gets the cost from those

then once its got the costs for each edge, it does the funnel algorithm based string pulling

[atomic coral]

can you point me in the direction of what code you're referring to?

[ocean wren]

goddam it.. gimme a sec 🙂

[shadow furnace]

Any built in ways to cull unreachable navmesh islands?

[ocean wren]

Not that I know of

All of this is within FPImplRecastNavMesh

[shadow furnace]

I guess I get to implement it

Weee

[atomic coral]

@shadow furnace are you wanting to remove disconnected mesh bits from the built data?

[ocean wren]

There might be a function on the Recast implementation.. there's a few functions that sound like it

Just don't know if any of them are made available to UE

This code is kinda crufty 🙂

[atomic coral]

hmm. there are facilities for disabling string pulling, but i don't see any way to Get There From Here

oh, there it is in the path following component

oh, nice, we already extend from CrowdFollowingComponent, so i can just add a new condition to disable string pull. maybe that'll help

nope

[ocean wren]

Hahaha.. classic unreal engine "maybe that'll help"... "nope" 🙂

Sweeney giveth and he taketh away

[shadow furnace]

At runtime ideally

[atomic coral]

i would expect if they aren't connected, they're not going to be considered anyway? at least, usually

[ocean wren]

I guess you can walk the navmesh and just alter the polyref's so they can be filtered

[atomic coral]

ok i am actually going to shelve all of this mess, because all of this work has had zero change whatsoever to the actual behavior in game. only minor changes in a test environment

[ocean wren]

Its a learning experience, get used to it

[atomic coral]

yeaaah i've been doing this for a number of years. i just hate every time i have to put something onto a shelf never to be seen again

[shadow furnace]

The issue here is we have a teleport ability that uses the navmesh

I suppose a simple "is reachable" check would suffice

[atomic coral]

ooooh

[shadow furnace]

Yeah, I should be more specific

[ocean wren]

Could you teleport and just look for the closest point on the navmesh next to the teleport target?

You can do a navmesh raycast for the direction of the teleport

[atomic coral]

i'm not domain familiar with EQS, but I'd try an EQS check with an is reachable test, rather than going through trying to eliminate disconnected bits before hand. That's probably the more reasonable method.
If you'd like, though, I have just figured out how to implement NavigationQueryFilter and you can use that to filter out

[celest python]

What does "search start/end" means for BT nodes?

What is a "search"

[misty wharf]

iirc "search" is when the node is active and the BT is "searching" for the next node to execute below it

so basically it should start every time the node becomes relevant and and end as soon as it stops

[celest python]

So like second state after "activation"?

[misty wharf]

Yeah something like that. Best to look at how the code is ran tbh

[celest python]

Gotcha, thanks

[celest python]

Is there any way to make BT wait for a node to abort itself before switching to another node?

Recently I started to think Abort function being synchorous is almost root of every problem I'm having in BTs

[misty wharf]

I vaguely recall looking into this as well but unfortunately the answer seems to be no

[celest python]

Damn

[misty wharf]

You could potentially have it set up so that you run abort logic in the controller or pawn, and have another node which waits for it to finish

or run abort logic separately in another node, or something

there's a few options like that I guess, depending on what it is you're aborting and how the rest of it works

[celest python]

I'm trying to make my AI switch to a specific behavior smoothly without cancelling everything instantly, but over the time I end up using FSMs for that rather than BT and I think I'm doing something wrong or BT's really not that efficient.. Probably I'm doing something wrong because whole industry using it

[misty wharf]

BT's aren't the answer to everything

[celest python]

I can also make sense aborts being sync because over the time conditions can change dramatically and you also have to manage the "aborting" state inside the tasks but.. 🤷‍♂️ It's ending up on choosing your poison

[celest python]

Over the time I start to approach them to more of a high-level layout of AI rather than a complete solution because 99% of the time I find myself needing to find a workaround for the issue I'm having and any workaround I find break the flow of BT

And over the time I start to feel like game AI is a very tacit knowledge

Like level/game design

Something you cant learn from books/articles/any other resource and only experience

[misty wharf]

Hard to say yeah

I think some type of patterns book would be helpful in theory

Eg. here's BT's, here's a bunch of things they're good for, here's a bunch of things they might not be good for

here's FSM's, etc.

[celest python]

Yeah, I wonder how AI people being recruited these days with this amount of lackiness of resources for those things

[misty wharf]

I never really looked at what kind of books and such are available for this kind of general purpose gamedev stuff... there's a lot for programming in general about just general purpose language/framework agnostic stuff

[celest python]

There is Game AI Pro 360 book, but I can say it's not really providing solution to real issues of the tools from what I read yet

But they show some real world examples from AAA games

[misty wharf]

http://www.gameaipro.com/ this is the one I know actually and seems to have some useful things

haven't gotten around to reading any of it tho lol

[celest python]

I strongly recommend it though 😄

Its very helpful

[lyric flint]

IS there a way to accomplish a "one time event"? So for example, I have a bool called CanSeePlayer and whenever my ai perception senses the player, I want it to do a one time trigger, for example play spotted audio and maybe a dialogue line, then reset and do it again next time the player is seen

[misty wharf]

Sure, just trigger it from when the player is spotted

successfully sensed will be false for events where LOS is lost so you can use that to check

perhaps also a timer to make sure it doesn't keep happening too often if sight is only lost for a moment

[mint terrace]

can I just say #gameplay-ai has been lit lately, love to see all the activity in here

[celest python]

#gameplay-ai is actually quite active though 🥲

[severe cypress]

Is there any system readily available that can do hundreds of AI without completely obliterating performance?

[ocean wren]

Mornin all..

There's some new GameAIPro style books being written too.. but won't be done until next year

[shadow furnace]

#mass is your best bet. Pretty spartan ai features but more on the way

Or roll your own

[viral elm]

is there any documentation for this?

otherwise I suppose I'm gonna have to read through these?

[misty wharf]

Usually the answer to "is there documentation for this" is "no" even for things that have existed in the engine for a long time

Sadly :P

[ocean wren]

That'd be a biiiiig NOOOOO 🙂

[viral elm]

😔

damn

Coming from unity and especially python, where everything has a documentation this is a change

[ocean wren]

You could always learn it all and write documentation for it..

python docus are terrible though, unity is better

But yeah, Epic have form in regards to poor documentation

[viral elm]

python docus aren't great - true.

But at least they are docus

[ocean wren]

Yeah.. can't argue

[misty wharf]

You can sort of go through the code usually to get the same level of docs as unity has

although unity docs do sometimes have at least a few samples

[ocean wren]

Yeah, at least they show you some examples of functions in use

[misty wharf]

for many engine bits you can search for usages to see how things should be used, but it might be harder for something like AI stuff :P

[viral elm]

it's intersting - because googling this plugin you find absolutely nothing

[ocean wren]

Yeah, if its blueprint and art related, you're golden

[misty wharf]

I think that plugin is almost brand new which is why

[ocean wren]

Well, its because it was a weekend project for one of the guys at Epic, hence the experimental tag

What is it you're after doing?

[viral elm]

still a shame you can't even find a mention of it anywhere 😅.

anyways, I'll try my best to read the code

[ocean wren]

There are plugins on github

For random pytorch stuff

and TF

[viral elm]

Trying to set up reinforcment learning

[ocean wren]

oh 🙂 ahahahaa... yeah no, doubt that you'll find good examples of that

Hoping a student of mine tackles that this coming year.. but haven't come across anything

[viral elm]

I've last checked a few months ago and none of them had support for UE5. which was understandable at the time

[ocean wren]

yeah, you'll always have to compile them yourself..

[viral elm]

yeah ... sadly.

With unity it was easy - MLAgents provides some great stuff

[ocean wren]

MLAgents was pretty crap too though 🙂

[viral elm]

i gues I didn't go enough in depth on it then 😅

[ocean wren]

One of my undergrad students used it for a project.. I mean it "worked" but it wasn't "nice"

As with many things in games, ML gets no love.. usability is kind of shocking

Plenty of work still to be done. With RL in particular

[misty wharf]

I wonder how hard would it to be to make your own weather AI 🤔

[ocean wren]

The guys at Embark rolled their own tech for it.. which is what I'm doing. Building a kubernetes cluster for training etc.

[misty wharf]

There's a website which shows rain radar stuff and you can usually predict if it's going to rain in the next hour or so by just looking at it

[viral elm]

I used it to set up some enemies for an arcade racer and it really worked nicely. But that' really all I've used it for. So not a a lot

[misty wharf]

the weather report usually just says "yeah it's gonna rain today" and not if it's gonna rain in the next 1h or so which would be more useful to know

[ocean wren]

We were doing level generation for this student project.. I mean it worked in a notional sense, but he was comparing different methods and some of them didn't seem to function

The aim was to look at learning by demonstration

Anyway, if you're just interested in RL then maybe that experimental thing might serve you.. haven't looked at it myself

But for production stuff, you'd be better off rolling your own kubernetes based training

kserve and all that

kserve->kubernetes->k3s->k3d

containerised python code, running in kserve on kubernetes cluster running on k3s in a k3d container 🙂

etc.

Basically, MLOps 🙂

[viral elm]

It's mostly really only for personal interest 👍

[ocean wren]

Let us hear how you get on

[viral elm]

although this does sound great 🤔

[ocean wren]

Its "the future" 🙂

[ocean wren]

goddam exploring Dall-E prompts is just too much 🙂

Trying to see if "Cyberpunk Zealot" overpowers "Studio Ghibli" 🙂

[misty wharf]

https://pbs.twimg.com/media/FUXYieMWAAA92Qn.jpg someone did this with dall-e mini lol

[ocean wren]

Apparently Studio Ghibli stands no chance against cyberpunk warrior 🙂

I am trying princess mononoke, but it might need the big guns.. totoro 🙂

[uneven stratus]

Does anyone know a good tutorial for making your own AI movement system that uses the navmesh? Characters are too expensive for my purposes and Floating pawn Movement don't give me enough control

[misty wharf]

You could just look at how CMC works and pick the bits you like

Basic character movement doesn't need much more than just a gravity force

[uneven stratus]

That would probably be a good place to start

Although I'm a little intimidated by all the complex code that works with replication (I don't need replication in this case)

[misty wharf]

Yeah I mean if you want to keep it as simple you can basically just have a movement force and a friction value

have it tick and it applies the force as movement on tick, and applies friction to slow it down if it's grounded

I implemented quake 3's movement code in unity at one point, their code is free and not too hard to follow

[uneven stratus]

Yeah that sounds like a good idea, I'll go take a look at that!

[stiff linden]

Okay, so I need some pointers of what I fucked up. I started of with the TopMap sample project. All was fine, the Navigation worked, Even an AI Move To node worked. But then at some point I changed something and now no ai movement at all works, except MoveToLocation but only when I turn off Use Pathfinding

My NavMesh is still there, visible by pressing "P"

Oops. I take it back, AI Move To does not work in that sample project

[patent hornet]

AIMoveTo requires an AAIController to work, as it needs its UPathFollowingComponent

players don't have those

[stiff linden]

Right. So I guess to summarize my problem: AI Move To or any of the move to functions don't work. The character does not move at all, no log outut, nothing

The class is a child of AIController

[ocean wren]

Did you try delecting the navmesh volume and readding it?

deleting even

[quick current]

In EQS, How do I put a custom circle center ? I want to put to the location of a Target Actor.

[misty wharf]

Create your own EQS context which returns the actor you want

[quick current]

Can it be modular ? Such I don't want to put 1 type of actor but a custom inputed actor ?

[misty wharf]

With some difficulty

EQS is not very easily parametrized

Your context can read the value from for example another actor or some other location to return it

[quick current]

Alright, if I understand correctly, it's going to be difficult to do this. My Target Actor is however present in a blackboard

[misty wharf]

One way which might not be too bad would be doing something like have your context get the querier

you can store the target you want in the querier, or you could even get the querier's controller and read it from its BB

[quick current]

Ok thanks a lot. I'm going to try this

BTW, I found a really good tutorial serie for AI, it's pretty advanced :
https://www.youtube.com/watch?v=6AiCpXOlrdE&list=PLNTm9yU0zou5y0cWEz59BQnHd0c6fC2XZ&index=26

[fossil wraith]

Hey everyone im making a AI path follow system but i am so stuck

This is the blueprint i am making its a mess and i think its all wrong but i dont know hwo to fix

This is the Function that im calling

void UPathFollowerComponent::GetNextWaypoint()
{
    if (CurrentWaypointIndex == Waypoints.Num() - 1)
        CurrentWaypointIndex = 0;
    else
        ++CurrentWaypointIndex;
}

[stiff linden]

Yep

[misty wharf]

You need to explain what the problem is because from what you said so far it's pretty hard to tell what even is the problem

[fossil wraith]

its not doing anything at all

[misty wharf]

so execute AI never runs?

[fossil wraith]

it is but i dont think I am getting "target" correctly

[misty wharf]

it's pretty hard to say because it isn't visible where it comes from

[fossil wraith]

the blue lines that are off screen is just another call of the targetvariable

Am i even doing the general code correctly?

[misty wharf]

I mean the basic idea looks correct

You need to debug it more than just "I think"

Use print nodes nodes to see what values are being set into the variables and where the code is actually going

[fossil wraith]

yeah i just tried and the Target is not getting set

but i dont know how to make a call to the component

[misty wharf]

Okay, so where does the target come from?

[fossil wraith]

its a reference to my path follower component because thats where the functions are

Should the functions really be in the waypoint actor?

this is what im trying to do for context

[misty wharf]

The ones that are shown in the screenshot probably not

so Target is UPathFollowerComponent?

[fossil wraith]

yes

[misty wharf]

Where does it get that?

[fossil wraith]

I dont think it does and i dont know how to

ive tried making a function that returns a pointer to the component but that didnt work

[misty wharf]

Right

The component is on your guard npc I take it?

[fossil wraith]

yes

[misty wharf]

In that case what you can do is cast controlled pawn into your guard type, then get it from that

or you can do a get component by class from controlled pawn

either way should work

[fossil wraith]

and the Behaviour tree this task is on the AI controller which is controlling theNPC

english failed me

[fossil wraith]

@misty wharf How do you suggest doing this because every single time i fix one thing another breaks so do you have a better way of doing this system

[misty wharf]

the basic idea is good you just need to debug it and fix it

[fossil wraith]

i know you cant see but im confusing myself like mad

OMG I COULD HAVE A FUNCTION TO DO THE CHECKS

isntead of spaget code

[misty wharf]

heh

having it as a function isn't going to make it work if it doesn't work at all tho lol

[fossil wraith]

i think because im using so much spagetti code it is breaking my brain therefore breaking the cod

[misty wharf]

:D

Did you get the casting working for getting the component?

[fossil wraith]

yes

The character moved to the first point

i just need to fix the looping

then i realized i wasnt incrementing the array

of waypoints

but my code was so messy i didnt know where to put it

[misty wharf]

Ah

It might make sense to set it up so that you have a function to which you pass the current waypoint

and it gives the next waypoint it should go to from that one

this way you don't need to keep track of the current index, and it will work regardless of what waypoint you start at

[fossil wraith]

already have that function 🙂 just need to make it so i can return the index

[misty wharf]

Wouldn't it make more sense if it returned the next waypoint directly instead of just the index?

[fossil wraith]

so make it return a waypoint actor instead?

[misty wharf]

so you can do something like move to waypoint -> get next waypoint -> move to waypoint...

[fossil wraith]

@misty wharfis it better that im using components rather than throwing everything in the AI character class and should i keep using it So i can make like chase component and like kill player etc as some enemys cant move and some cant kill

[misty wharf]

If it makes sense to you, sure

[fossil wraith]

well as im going to have multiple types of guards i would have thought it would have been a better workflow

[misty wharf]

It's a bit hard to say whether it's the correct way or not, because there are many ways of doing these things, and whether it's good depends on how you want things to work

[fossil wraith]

im mainly worried about optimization and performance too

which i guess i can profile that at a later date

[misty wharf]

Actor Components should be fairly good. Scene Components will affect the performance of moving things as they need their transforms updated when moving

But in general unless you have hundreds and hundreds of characters moving about, you're unlikely to see scene components being a problem

[fossil wraith]

I dont plan on having massive levels (I suck at level design and i am still thinking of how i want my levels to be

but my plan is for my game is having an angel thats fallen from heaven by accident and has to sneak their way back to heaven and humans hate angels so yeah

[fossil wraith]

Apparently it cannot get to the waypoint

Just worke dout why its because every time my task runs its reseting the Array

Okay so i need to have the array on the NPC?

How should i se the initiall variables should i have like a manager of some sort?

because its resetting every time the code executes

@misty wharf sorry to be a pain again what should i do here?

[misty wharf]

You probably should store the current waypoint somewhere else than the task

that way you can then read the current waypoint in the task to select the next one

[fossil wraith]

is there like a type of class or soemthing i cna store it in or where else could i?

or do i just make a function in the component that is like "get current waypoint"

thats probably the best bet?

[misty wharf]

Yeah you can just store it where ever it makes sense, could be in BB, could be in ai controller, or in the component

[fossil wraith]

black board would probably help in the future when say i do my chase stuff when it returns to normal?

however i think quite a few of the variables only need to be set once at the start then just updated

[normal mist]

Does anyone know what a "binary broad-phase filter" is?
context (timestamped): https://youtu.be/5ZXfDFb4dzc?t=2885

[celest python]

Probably zoombapup might know that, but I really liked that yt channel

[ocean wren]

He just means a boolean condition that seperates the logic into true side and false side

So filtering in this case would just check if true and enable specific behaviours to check if that were the case

imagine something like having ammo being a filter for all of your weapon handling logic

So its a boolean selector in a behaviour tree.. or a node in a decision tree, or a conditional value in a utility system

The "broad phase" thing just means at the topmost logic level.. before you get down into the details

You'll often hear of broad phase and narrow phase.. typically in physics, where you do a sort of higher level object sorting as broad phase, then individual overlap tests as narrow phase

[normal mist]

wow, thanks a ton @ocean wren

[gritty peak]

Hey I am having an issue with my ai. When i spawn them in have a them patrol on a lane. For some reason certain enemy types do this fine where others get their navigation aborted. The odd thing is it only happens on one level. any idea why this might be happening?

[bronze wagon]

My behaviour tree is crashing the editor and I cannot figure out why. It’s been happening since last night and I lost several hours of work due to it. I spent today setting up perforce and making commits before each run now. I’ve gone over the blueprints over and over but cannot find the issue.

First it was a memory exception and now it is an array out of bounds exception

I’m using arrays but only for each loops

Not sure why it would crash the editor instead of the PIE and give me a hint on what is happening.

Seems random but it doesn’t take long to fail

Crash output

[bronze wagon]

I just stepped through every line in the code and have a breakpoint at the begin of every task, no issues. yet it crashed after the node was complete for the first time. it is a looping branch

looking at the unreal source it looks like it is the debugger itself crashing

[ocean wren]

It could be that you're stomping memory and its just appearing in the BT stuff. Try and isolate just the bit of functionality. So maybe copy your BT into an empty level with just what you need to make the BT function and debug it there.

Essentially, you need to make a reproduction case before you can fully debug stuff like this

Also, consider what assumptions you've made, check if they are true.

[keen crow]

why does the pathfinding test discard actors that are reachable? is it something about the fact the actors root location is in the navmesh gap?

[keen crow]

it seems it has to do something with these properties of navmesh. I tried lowering them for some other actors and it broke pathfinding for another actor...

[mint terrace]

well if the actor is in a gap between navmesh then it's not reachable via navmesh

I'll be here all night

[bronze wagon]

Well I don't understand why but apparently I needed to move the looping call to the selector at the top of the graph vs the branch I wanted to run until a BB key changed. Really it didn't need to be that far down, makes sense to have it at the top since all of my branches are state specific

no crashes yet

[keen crow]

well the gap is generated by the actors static mesh. I could of course disable "Can Ever Affect Navigation" property but that's not what I want

[misty wharf]

I've had this same issue

I solved it by just having it check path to front of the object instead of to the object's location

Of course for me that works because the objects have a clear "front" which must be reachable for it to work correctly, but for something else it might not I guess - I'd imagine there should be some kind of way to do a custom reachability check where you can test distance to the final location

[mint terrace]

well if they were modifiers as opposed to nav mesh occluding then it would still be navigable

i mean---- if they are not supposed to move, but also do not block navigation, then Can Ever Affect Navigation set to false might be the right call

if the actor is supposed to move, and you're doing dynamic navmesh generation, then yeah thats a pickle

[misty wharf]

I don't think not having it affect navigation is the solution though

If it doesn't affect navigation, npc's would try navigate through it which would obviously not work

[mint terrace]

I don't know what the target is for, I guess

[misty wharf]

Yeah I mean it would in theory work for that, but for example if an NPC needs to walk past it, it wouldn't be able to avoid it being there if it doesn't affect navigation

[mint terrace]

yeah assuming it's an actual thing to be avoided

/shrug

looking back at that screenshot I'm not sure it is something to be avoided?

more like a destination for ai to take cover or something (red circled part)

[misty wharf]

Well yeah, but even if it's intended as something like that, you'd walk up to it, and then if you'd decide you need to walk to something on the other side of it, you wouldn't be able to because you'd get stuck on it if navigation was disabled on it :)

[mint terrace]

if you disabled affects nav the navmesh could generate beneath

[misty wharf]

yeah but if you tried to move to something which is on the opposite side of it and navmesh was generated under it, the npc would just attempt to walk through it

[mint terrace]

yeah that's true

there are reasons, but then I'd wonder if making it higher cost would be better

then navs through it could be controlled, but navs to it's point would be valid

[misty wharf]

Good point, I guess as long as its movement cost was sufficiently large that it would never make sense for an npc to attempt to "shortcut" through it it might work

[mint terrace]

yeah tuning would be needed there

but if you are naving directly into a high cost area, then no other path would suffice anyway

would always be higher cost

so whatever cost necessary would work, assuming this didn't kill other pathing flows

if it's a cover spot, most times you don't want any AI transiently going in there or else face awkward RVO avoidance

or worse, detour crowds 🥁

[strange blade]

Hey anyone played much with generating splines? I'm trying to get my ai to do some 'dinosaur movement' behaviour, and get them to follow a spline! Unfortunately the spline I generate from path points along a navmesh suck and was wondering if anyone had any suggestions to be able to get a more smooth transition?
Repost from #blueprint but was told it might be a good question here? Also worth mentioning I'm trying to achieve this with blueprints hehe

[mint terrace]

ok so

you're selecting some points along the spline to generate?

what if you produced the point on the spline to generate over time with a timer or a relaxed tick, and then fed that into blackboard with observe blackboard value

youd be able to control the resolution of the calculation, and observe blackboard value woudl allow moveto's to the point to update on the fly

should allow you to control the tradeoff of cpu vs accuracy of moveto along the spline

should be plenty doable in blueprint

[strange blade]

Interesting, im looking into a way to produce this now. I'm currently using 'find path to location synchronously' and generating my spline based off those points. Would it also be possible to get the length of that spline and generate say 10 points across that spline with point 0 at the root, 9 at the final length, and then spread the remainder across the rest with equal distance apart?

[mint terrace]

if blackboard doesnt work for you then direct find paths are obviously needed, but telling the AI to just navigate to a position with blackboard/behaviortree might be what you are looking for [using observe blackboard to update pending move-to's on the fly]

thats the most obvious path. if you are already invested on direct findpath calls, then maybe use the async version and interrupt it as the up to date point along the spline changes

im suggesting ticking at a low rate to select these points along the spline as the current game time increases

[strange blade]

Thanks I'll investigate the behav tree approach! I'm still quite fresh to using them so I have a bit of a learning curve to understanding good practices and such.

[mint terrace]

you can do a lot with it; its easy to think that its janky because moveto has to complete to the original position you requested and that you cant change that, but you can

[strange blade]

will that moveto just be updated when the blackboard position changes?

[mint terrace]

if you have ticked "Observe Blackboard" then yes

[strange blade]

Interesting! I'll check that out-

[strange blade]

Also with this, what do you suggest I should use to generate the spline point?

[mint terrace]

generate the spline points you mean?

[strange blade]

yes!

[mint terrace]

that part is the hard part 🙂

some code or some such to do it 🙂

[strange blade]

ha, I thought as much :p

[mint terrace]

recently we discussed a way to override how navigation costs worked at a core level to do something similar to this (limiting the required angular change of navigation paths to be minimized)

but that is pretty intense, and since you are intending to go for splines to generate the path ahead of time, i feel like it can be done without that level of difficulty

[strange blade]

I see, I was hoping that this way would get what I wanted (or close to it) without going through too many hurdles

[mint terrace]

there's also something to be said for just modifying the maximum rotation on the character movement controller

i think in theory its supposed to somewhat respect it

ymmv

[strange blade]

I was able to achieve the effect I wanted with 'add movement input' nodes to move if pawn see's player, and if not rotate and move until it directly sees player again, but sadly it had no accountability for path finding so that wasnt too helpful in the end :p

[strange blade]

I guess to a degree this works, but i found it less than desirable on anything that isnt humanoid or bipedal :\

Trying to move this fella becomes a bit more obvious sadly

But thank you for the help and pointers! It's definitely pushed me towards a better direction! I'll report back if I get anywhere with this! 🙂

[mint terrace]

yeah it's not easy getting cmc to mesh with realistic simple rotations

[fossil wraith]

AI is torture ISTG

[keen crow]

Actually worked for me, thx. There's even a NavModifier component

[severe tangle]

Can I somehow inherit from DetourCrowdAIController, because of the avoidance logic already implemented, and use it inside my own AI Controller?

[misty wharf]

It's not exported so no, but if you look at how it's implemented it's extremely simple and you can just do the same thing in your own AI controller subclass

<@&213101288538374145>

It's been a while since I've seen one of these, used to be a lot more common it feels like or maybe I've just not been looking when they get posted lol

[fair zealot]

:no_entry_sign: tagamedeveloper#8802 was banned.

[night hazel]

yeah

slight uptick recently

but it has been far fewer than they used to happen

[misty wharf]

It's good to know Discord is finally maybe doing something about it

[fossil wraith]

Zomg

[misty wharf]

Oh boy

[fossil wraith]

<@&213101288538374145>

[fair zealot]

:no_entry_sign: blob#4362 was banned.

[fossil wraith]

My gawd

[night hazel]

i just tabbed away

[misty wharf]

lol

[night hazel]

and it happened again

[misty wharf]

they're back with a vengeance

[fossil wraith]

Zomg do you know why my location isnt updating and is this even correct?

[misty wharf]

Is it supposed to update the BB value or?

[fossil wraith]

yes and i just realized thats the old code

<@&213101288538374145>

[misty wharf]

wow

so much for discord doing something about it lol

[night hazel]

yeah

looks like they figured out how to circumvent the blocking

[fossil wraith]

jesus between the neo nazis raiding servers supporting pride and this discord is a mess

IVE FUCKING DONE IT

[weary holly]

Im trying to coordinate my ai and I saw that there is a team sense. Now I've found only some articles about how to use them but not really in blueprint, could somebody explain to me how the sense exactly works and what stimuli information it returns?

[misty wharf]

I don't think anyone has really used the team sense, it's come up a few times but it seems nobody really knows

In general the builtin team support with the attitudes hostile/friendly/neutral is something you need to use C++ to configure

[opal crest]

The heavy duty tool for this is probably to customize the path following component, and control how your agent moves toward the next goal point on each tick. I'd guess that the faster your dino is going the slower it should turn.

https://docs.unrealengine.com/4.27/en-US/API/Runtime/AIModule/Navigation/UPathFollowingComponent/

Rama has some code/videos using this to add jumping to the path following behaviour of his AI.

The other thing you could do is manipulate the generated spline control points. But I honestly don't have the a good clue how you would wrangle to the math to get the intended effect (Maybe Freya's Bezier videos on you tube would help there?). I use some very simple splines generated from navpaths to visualize turn based movement, but for those I make the handles really short to make the corners tight but rounded.

[strange blade]

Thanks for getting back! 😄

[severe tangle]

I'm playing around with DetourCrowdAIController and NPCs are avoiding each other as expected, but it seems they ignore player's pawn. Can I change it to avoid player's pawn too? Using Blueprints

[misty wharf]

You need to add the player pawn into the crowd manager

[ocean wren]

You can Implement the DetourCrowdFollowingComponent as your pathfollowingcomponent

[severe tangle]

Crowd Manager can be only changed via C++? Don't see a way of adding players pawn inside Project Settings => Crowd Manager.

[severe tangle]

Sorry but how could I achieve that?

Oh if you meant CrowdFollowingComponent in Path Following Component I have already done that, but it still ignores player

[misty wharf]

iirc yeah the crowd manager might be another C++ only thing

but it's basically CrowdManager::Get(World)->RegisterPawn(player pawn) so it's not super hard to do

[severe tangle]

oh yea that should be doable! Thank you

[severe tangle]

Sorry to bother, but I tried your approach, looked up UE docs as well as forum post

https://docs.unrealengine.com/5.0/en-US/API/Runtime/AIModule/Navigation/FCrowdAvoidanceConfig/
https://forums.unrealengine.com/t/registering-player-character-with-ucrowdmanager/313220/3

My Implementation looks like this:
``
UCLASS( ClassGroup=(Custom), meta=(BlueprintSpawnableComponent) )
class MYPROJECT_API UUCharacterCrowdAgentInterface : public UActorComponent, public ICrowdAgentInterface
{
GENERATED_BODY()

public:
void InitializeComponent();
// Sets default values for this component's properties
UUCharacterCrowdAgentInterface();
};

void UUCharacterCrowdAgentInterface::InitializeComponent()
{
Super::InitializeComponent();
UCrowdManager* CrowdManager = UCrowdManager::GetCurrent(this);
if (CrowdManager)
{
ICrowdAgentInterface* IAgent = Cast<ICrowdAgentInterface>(this);
CrowdManager->RegisterAgent(IAgent);
}
}
``
I also added this component to player via constructor, but nothing seems to be happening.. Do you know why? Or how could I make it work, I wasn't able to find some decent tutorial sadly..

[misty wharf]

Ah right it required that interface

You need to implement the functions in it - GetCrowdAgentLocation, Velocity, Collisions and MaxSpeed

You can just have it return the owning actor's location, velocity, capsule size and radius and max speed respectively

[severe tangle]

I see, I thought I could just add it as a parent from UE editor but this interface wasn't present. Going to give it a try

[severe tangle]

I did implement them:

void UUCharacterCrowdAgentInterface::InitializeComponent()
{
    Super::InitializeComponent();
    UCrowdManager* CrowdManager = UCrowdManager::GetCurrent(this);
    if (CrowdManager)
    {
        ICrowdAgentInterface* IAgent = Cast<ICrowdAgentInterface>(this);
        CrowdManager->RegisterAgent(IAgent);
    }
}
FVector UUCharacterCrowdAgentInterface::GetCrowdAgentLocation() const
{
    return GetOwner() ? GetOwner()->GetActorLocation() : FVector::ZeroVector;
}
FVector UUCharacterCrowdAgentInterface::GetCrowdAgentVelocity() const
{
    return GetOwner() ? GetOwner()->GetVelocity() : FVector::ZeroVector;
}
void UUCharacterCrowdAgentInterface::GetCrowdAgentCollisions(float& CylinderRadius, float& CylinderHalfHeight) const
{
    if (GetOwner()) {
        GetOwner()->GetSimpleCollisionCylinder(CylinderRadius, CylinderHalfHeight);
    }
}
float UUCharacterCrowdAgentInterface::GetCrowdAgentMaxSpeed() const
{
    return 300;
}

But result didn't change

[misty wharf]

Maybe try registering in BeginPlay

In my other project where I use this, I have it in my player character and it registers itself in BeginPlay, the implementation of the interface is otherwise pretty much the same as yours

Also you shouldn't need that cast to interface there, since the class inherits from it already

[sudden citrus]

Who do I ask if there is dynamic subtree injection planned for the State Tree?

cause if there is I can safely throw BTs in the bin

alternatively: please make blackboards stackable

then I won't have to throw them in the bin

[celest python]

They were talking about Epic was going to listen the community at #mass

You can also ~~hijack ~~ sneak StateTree into their scope @sudden citrus

[sudden citrus]

Can you explain what you mean by that?

[celest python]

Ehh, I hope my English didnt fail me again 😂 AI team of Epic told they are going to listen community to add new features

So you have a chance to provide them these ideas

Ask someone at #mass

[sudden citrus]

I mean the hijacking

[celest python]

Ah that was totally an incorrect word usage of mine

[sudden citrus]

oooooooh you mean mentioning state tree in that topic, I thought you were talking about a technical solution 😂

[celest python]

lol yeah

[sudden citrus]

I doubt dynamic injection would be possible in the context of Mass at least - it would change entity memory layout

I'm hoping to be able to support it in a Brain Component though - simple FSMs for crowd members and low LOD actors, complex in combat

[celest python]

Since StateTree is used together with mass maybe they can also prefer to get feedback for those seperately

That's what I tried to say by hijack sneak into their scope

[sudden citrus]

let's see, I'll have a look through the channel, thanks

[celest python]

For now they'll only listen community for mass specific things but state tree is also pretty much related

[misty wharf]

I wouldn't hold my breath on them listening to anything

Nothing I've seen them promise they would do on the forums has ever materialized

[sudden citrus]

I have slightly higher hopes for UE5 features seeing how they've hoovered tech talent throughout 4's life

probably backwards compatibility standing in the way of change more than anything

[misty wharf]

If you want a feature in UE you need to convince Fortnite players it's the best thing ever so that they all start begging for it

[final trail]

I have a melee enemy and I’m trying to figure out how to make them attack at intervals instead of at the same time, any tips?

[misty wharf]

Easiest solution would be to add random delay between actions

this way they will not synchronize with each other even if they begin actions at the same time

[final trail]

I see

That’s way easier to set up than a director

[misty wharf]

Yeah, it really depends on how you need it to work but if all you need is for them to not look like they're acting in perfect sync, random delays between actions should do the trick :)

[final trail]

Thanks so much for the input

[placid bloom]

Hey all!

Noob question here but for an AI just chasing a player is it better for optimization to use the blackboard or just a MoveTo off and event tick?

[celest python]

for a single AI it wouldnt matter

but to answer the question using MoveTo node in BT is more optimized

Only path following component ticks in C++ to move the pawn and MoveTo node only listens for movement end or failure events etc

[placid bloom]

ok fantastic. I was getting to about 15 ai and it was starting to chug a bit so ill look into swapping it over to that.

[celest python]

always profile first

bottleneck might be something totally different

moving 15 ai should be nothing

Like only should cost 0.2ms or even less

[placid bloom]

well they are only are only 2d flipbooks so they shouldnt be the issue

[celest python]

Check this link: https://docs.unrealengine.com/4.27/en-US/TestingAndOptimization/PerformanceAndProfiling

[placid bloom]

ok ill have a look thanks 🙂

ok so did some testing and its definitely doing it only when the bot is moving. This is the setup i was using

ok i think it might have been the part at the end causing a loop even though it has event tick. Not really sure why i had that in there

[celest python]

thats not correct though

AI Move To already ticks behind the scenes, you dont need to update it with your own tick

AI Move To is pretty much same with BT's MoveTo task

[placid bloom]

so just have it go off event begin play?

[celest python]

yeah

also try to use BT instead

unless you really need something basic like this (which is fine)

if your game mechanics will extend on the long run its better to setup a BT

[placid bloom]

yea im just looking for something basic. Their aim is to get to the player (Which will deal damage) until there is no health left. Do i have to do the custom event to reset it because they seem to stop as soon as they reach me

[celest python]

probably MovementResults returns "AlreadyAtGoal"

cover all the cases and tie them to Chase node

from OnSuccess and OnFail outputs

by using Movement Result enum

[placid bloom]

Ok sweet that has fixed it. Now i can have over 100 with no change to the fps

[celest python]

Nice

[placid bloom]

Thanks for the help 🙂

[celest python]

Np

[noble silo]

Hello! Has anyone had much experience with using AI in a top down paper sprite game? My main issue is I am using sight as my main source of perception but my enemy can always see my main character through the walls because they are almost horizontal. Would sound be a better sense or is there a way to make AI not see through collision?

perspective on my ground plane

game view

[misty wharf]

it uses visibility as the trace channel by default

so as long as you block on that channel f.ex. via invisible colliders it should work

[noble silo]

ok cheers!

I will look into that!

[misty wharf]

depending on how things work you can also change how the sight check trace is done, but for that you'll need C++ to implement IAISightTargetInterface

so you could for example trace closer to the ground to ensure it hits the low walls

[ocean wren]

I'd just build blocking volumes for my tiles

I assume you're using tiles

Hmm, I might have messed up.. I agreed to do a quick visualization simulation thingy for a navy event... probably going to cause me pain 🙂

I guess I could stick in a T72 or two and call it integrated intelligence or somethin

They want a visualization of some sattelites.. so might have to draw some satellite orbits

Its just a quad strip though thankfully..

I guess I'll have to write a scanning shader so we can show some ships being scanned by the sattelites