gameplay-ai

[misty wharf]

Eg. NavArea_Slime or whatever

Then you can create a Nav Query Filter for your NPC's which has a higher cost on NavArea_Slime

you can then set any npc that needs to avoid the slime to use that particular query filter

[keen spade]

Thanks for the answer! You said the key part I needed to be been searching for, Nav Query Filtering, not looking around to change nav_obstacle

I found this great video for anyone interested https://www.youtube.com/watch?v=dDnHAUgaQQk

[hallow flint]

ohhh tell me more! Any particular talks that caught your eye?

[ocean wren]

Well, a bunch of the neural radiance field papers are pretty interesting, I've got a MSc student looking at NeRF vs Photogrammetry failure conditions, so had a dig through those

There were like a few hundred papers on NERF from CVPR

the other thing is the new text->image generators like DallE2 and Imagen, those techniques are being applied to things like 3D mesh generation and Character generation

still really clumsy and shit, but I can see the direction of travel

Basic idea is pretty cool, you type in a text description and it spits out a 3D mesh generated from the text latent embedding

[celest python]

A few weeks ago I met with a professional character artist, he told me some artist are anxious about not just humans but also creatures are going to be generated by AIs in the future.. Do you think that techniques are eligible enough to create a witcher monster for example?

[ocean wren]

Personally, I can see the case for NeRF models being used alongside the CLIP guidance style stuff, then remeshed (because NERF's are just mathematical probabilities they don't have resolution)

[celest python]

I can see existence of metahumans depressed some of the artists

[ocean wren]

Yeah, honestly, the key is having some data to train on.. if you can think of somewhere we can capture multi-view images of witcher monster like things.. then we can NERF them and generate them

So for instance, video clips of witcher monsters 🙂

or hell, just proper sculptures of witcher monsters captured as nerf models

all of this ML stuff just requires some way of having loads of data to train on.. if you can think of ingenious methods of doing that, you're onto a winner

all these text->image generators are just trained on large text->image pair datasets (470 million in Dalle2 if I recall), there's a public dataset called LAION5B with 5 billion images in now

[hallow flint]

ahh that's co cool

So when it comes to 3d models

[celest python]

I'm surprised with that amount of innovation, someone hasn't created a auto-retopology AI yet

[hallow flint]

there are tons of resources for getting them, what am I missing here?

[ocean wren]

Eren: there are 🙂 just not great yet

autopo stuff in 3DCoat is actually pretty good, but not great

[hallow flint]

there are tons of websites of existing assets, couldn't all of those resources be used for training data?

[ocean wren]

Josue: copyright 🙂

[hallow flint]

😮

i think google released some site with models that weren't too strict on copyrights

[celest python]

I was also thinking of InstaLOD, but none of them able to retopology a hard surface model

[ocean wren]

Honestly, I imagine companies like Quixel are going to dominate the market for datasets of 3D meshes

[celest python]

Only organic ones

like apple or human face etc

[hallow flint]

right

[ocean wren]

Its more likely that what'll happen is that some company like apple, that does 3D scanning, will actually use the data from the scanning to create a large dataset internally

did you guys see apple room scan?

[hallow flint]

i have not

[celest python]

not yet

[ocean wren]

one sec.

https://www.youtube.com/watch?v=9KcFNZ5Na64

[hallow flint]

crazy

that's super useful

[ocean wren]

Yeah, imagine a huuuge database of room layouts, with semantic tags for all of those objects, plus the dimensions and positions of the bounding boxes

[hallow flint]

i wonder how this works

visual-based lidar?

[ocean wren]

now... here's the thing... imagine that room plan scan.. where you can use language prompts to change the content of the room, but maintain coherence of layout?

i.e. "make all my furniture made out of wool"

"replace all of my carpets with neon floors"

the apple stuff is pretty standard computer vision stuff

[hallow flint]

yeah that'd be super neat

[ocean wren]

the main thing here is how stable the bounds are.. its super clean

they're using the lidar to guide the bounds generator

really great idea

[hallow flint]

actual lidar?

do IPads come with lidars now?

[ocean wren]

yeah, the iphones and ipads have a lidar scanner on

[hallow flint]

huh

didn't know that

[ocean wren]

since the iphone X

[hallow flint]

i don't really use apple products so this is news to me haha

[ocean wren]

yeah, I avoid apple where I can, but they do have the money for some cool tech 🙂

[hallow flint]

haha yeah that's true, im not a fan of how they go about doing things but

[ocean wren]

Speaking of evil companies, meta's next headset is looking pretty damn nice too 🙂

[hallow flint]

i gotta check that out

this?

[ocean wren]

cambria yeah

[hallow flint]

do you know anything about the current state of BCI?

last I heard was that Valve was putting tons of money on researching that

but nothing else aside from that

but that's not my field, i just find it interesting

[ocean wren]

We had someone working on BCI, well, sort of.. there's a device which I can't remember the name of.. will come to me in a sec.. it measures brain blood flow

basically shines a strong IR light into the head, then captures that light with a set of sensors and builds a map of the blood flow

[hallow flint]

ohhh brain blood flow, that's different than the typical EEG devices huh

was that device expensive?

[ocean wren]

yeah, dammit, I can't remember the acronym it was like FMRI but clearly not that

yeah, device was expensive

our "cheap" one was 15k 🙂

proper full head ones are like 100k +

[opal crest]

fMRI is something that UX researchers were playing with (functional MRI) about 10 years back. It does what you describe (measure blood flow in the brain to deduce which parts of the brain are active).

[ocean wren]

FNIRS!

[hallow flint]

damn

[ocean wren]

thats the one 🙂 goddam it.. knew it began with an F 🙂

functional near infrared spectroscopy

[hallow flint]

i bought a cheap eeg device

it was like 700

but turns out

if you have hair

they suck

haha

[ocean wren]

yeah, I had one of those too 🙂

[hallow flint]

so i haven't been able to use it

[ocean wren]

they were shit

[hallow flint]

i know, what a waste

[ocean wren]

basically, they captured muscle movement in the eyelid area more than the EEG signal

totally useless

[opal crest]

But you could play brainball 🙂 https://www.youtube.com/watch?v=7I1RFJnOaTY

[hallow flint]

😮

that's so cool

[ocean wren]

proper EEG's have some pretty impressive sensors.. and you have to use contact gels

[hallow flint]

haha

[opal crest]

Out-relax your friends 🙂

[hallow flint]

have you guys tried out the OpenBCI devices?

they're super cheap

but i assume

sensors are shit

maybe if i shave my head

i could get some decent results

lol

[ocean wren]

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

Ours is a bit like that

Not really my thing though.. the researcher who was using it was also hairless 🙂

This is more my thing 🙂

[opal crest]

I got my Mid Journey beta invite today.

[ocean wren]

I've got a thing for "brutalist megastructures" today 🙂

Ah nice one Epi!

[hallow flint]

lmaooo, dedicated reseracher!

[hallow flint]

ahh i got mine yesterday!

[ocean wren]

I'm just setting up some code to do my own diffusion stuff.. so I don't have to use external services or google colab

[hallow flint]

this is with dall-e or midjourney?

[ocean wren]

those images were generated using disco diffusion

there's also a thing called centipede diffusion I'm going to explore over the weekend too

[celest python]

How come those image generating AIs become a trend in like.. almost one day?

A few months ago no one was talking about them

[ocean wren]

More people are getting access

[hallow flint]

this is amazing

[ocean wren]

and a few of them are commercial services now

[lofty sail]

Hello, do somebody know why is this navmesh so inaccurate?

[ocean wren]

because you didn't set it up right?

there's a cell size on the recastnavmesh that you can tweak for instance

but of course a finer navmesh will mean slower pathing

Also the character size will impact where the navmesh is.. it won't generate close than the agent width, otherwise your agents would intersect with geometry

Good one from today

[hallow flint]

what was the prompt?

[ocean wren]

I dunno 🙂 didn't save it for those! 🙂

[hallow flint]

hahaha

[ocean wren]

it definitely had science fiction market 🙂

pretty sure it had brutalist megastructures in it too

[hallow flint]

i need to add this as part of my workflow when im game jamming

[ocean wren]

I'm going to try and experiment by posting some DallE images on IndieDB see if people spot them as not typical "concept art"

[hallow flint]

you could 100% fool me

[opal crest]

God, I would love to see a pinterest thing that extrapolates from my pinboard.

[hallow flint]

I would be there commenting 'what's your artstation link'

[ocean wren]

disco allows you to feed it a seed image.. so I'm going to disco-ize some DallE2 prompts 🙂

[lofty sail]

I played with the cell size for like 2-3 hours and I don't know what I did now, but it's working how it should, thanks!

[ocean wren]

🙂

Thats the way with game AI 🙂

I've actually started using generated images in my project pitching workflow.. so far has been super helpful

[celest python]

@ocean wren you never talked about Google AI becoming self aware, yet 😄 I'm getting suspicious 😂

[ocean wren]

two small funded projects using AI artwork to get the concept across

oh you say that? the general AI thing?

I'm really not that interested in that media bollocks 🙂

ML isn't intelligence.. its a maths trick 🙂

its what a load of ML people call a "function approximator" which is pretty damn accurate

[celest python]

when this is become something so normal

[ocean wren]

for any function F, if we can feed it examples of f(X) for enough X's we can calculate an approximate F

thats all it is.. an arbitrarily complex F 🙂

just a really cool useful F 🙂

[celest python]

what is intelligence anyway? it's nothing more than some neurons communicate each other mathematically..

[ocean wren]

chemicals, biologicals, electricals

and a shedload of quantum effects

[celest python]

but since all of them can be explained functionally, and with values, wouldnt that provide same result?

[ocean wren]

its kind of an insult to intelligence to suggest that a few matrix multiplies is ever going to happen upon that 🙂

theoretically, if you had infinite compute, then possibly.. but practically? not until we have fleshputers 🙂

[celest python]

fine, fair 😄

[ocean wren]

I mean quantum computers might edge us a tiny bit closer.. in like 100 years

but this whole google guy thinking a language model is sentient thing is total PR bullshit

[hallow flint]

haha they actually made that claim??

[ocean wren]

I don't know why people can't just be amazed at the things that are possible, or very nearly possible now you know?

[hallow flint]

wow

[ocean wren]

Yeah, but it was probably more PR spin than actual story

some guy was testing a language model for hatespeech and he basically led it to give him answers that sounded human

because its a language model and it was trained to do that 🙂

[misty wharf]

I don't know if it's PR, weirdly organized if it was... I think the guy who made that claim might just be a bit nuts lol

[ocean wren]

it was literally trained to be helpful and friendly, so it had the vibe and this guy thought it was sentient and started asking google's bosses to grant it rights

[hallow flint]

hahaha

[opal crest]

That guy was an AI Ethicist, and his specific claim was that because the bot claimed to be human (edit: a person), that claim needed to be taken seriously.

[ocean wren]

if you read some of his discussion text, it was obviously leading

[opal crest]

I honestly think it was a bit of a "I'm a very specific kind of researcher and I want people to pay attention to my niche" thing.

[ocean wren]

quite possible yeah

[celest python]

I dont think people like reality at all 😄 I think thats pretty much it

I dont think living in an information era makes people happy either

[ocean wren]

I mean, it wouldn't normally get traction, but the media and state of shit right now kind of lend it that "me too" thing

[celest python]

We can see AIs getting more benefit of it more than humans

[ocean wren]

I dont know, I love some of this AI tech

[opal crest]

Google didn't help themselves by Streisanding the whole thing by suspending him.

[ocean wren]

I just want to use it for game production tasks 🙂

I mean there's tons of issues with it ethically.. but dammit, I can haz pretty pictures! 🙂

The question I keep asking myself, is why the games industry is asleep at the wheel with all of this

[opal crest]

Day job me is excited by the possibllity of being able to allow for near professional quality, semi automated educational content creation. Hobby me would like to be able to have 'good enough' assets on demand.

[ocean wren]

I mean look at how poorly they treat mieszko 🙂

[celest python]

Especially Epic

[misty wharf]

Tbh I'd be really curious to see ML based layout systems

[ocean wren]

On the way 🙂

[misty wharf]

We have a tool at work which is used to do layouts and pretty often certain elements repeat in them, eg. image on left, call to action in bottom right

Having it automatically suggest a layout for you would be like a 10/10 groundbreaking feature

or even automatically derive a layout from an existing layout for a different size

[ocean wren]

my summer is being spent doing UE based ML generative layouts

[opal crest]

The iOS layout engine is based on a oonstraint solver, that due to runtime requirements, has a number of boundry issues. A fast ML based solution could revolutionize that part of UI design.

[misty wharf]

Yeah we're an ad tech company, we have a tool for creating what is called "responsive ads", basically ads which can be displayed in any number of sizes or shapes

[ocean wren]

oh you mean UI layouts?

[misty wharf]

so you can imagine a large portion of time for a designer goes into making sure the ad looks nice in various sizes, and laying them out

being able to get that even partially automated would be such a big deal

[ocean wren]

oh yeah, that is DEFINITELY going ML

[opal crest]

Not ad-tech, but I do torture CSS for a living sometimes, so yeah, that part of the job would be wonderful to automate.

[ocean wren]

especially website stuff.. because you can scrape the living crap out of the web and nobody cares 🙂

[misty wharf]

https://3k0ezklsacdme1qf1g16lf1b-wpengine.netdna-ssl.com/wp-content/uploads/2021/11/Hero-BG-16181B.mp4 I wonder if this will embed correctly

Okay it does

[ocean wren]

I mean imagine you just visit all the URL's with some poor people saying if they like the layout or not.. the ML trains on billions of webpages with different layouts.. learns the probability of that rating from the humans.. and bingo!

[misty wharf]

Yeah if we could automate parts of that that'd be a very big deal lol

[opal crest]

Games eventually warmed to houdini for using PCG to assist content creation. I see a similar path for the rest of this in games.

[ocean wren]

Very much so

[ocean wren]

Nice! I might even have a MSc student who could do that.. I'll ask her

[misty wharf]

Huh

If you know someone we might be able to do a Thing™️

[hallow flint]

So don't quote me on this BUT i do believe that adservices provides already have data that could be used as annotation

[atomic coral]

hey all 🙂 got a hopefully much simpler one than the last item that absolutely stumped me. I've got to be missing something.

I've got an AI that performs an attack while it is doing a MoveToActor call on the player.

On a successful hit, I call StopMovement() in it's controller, and then wait.

The character unexpectedly continues stumbling forward (and NOT towards the previous MoveToActor target) and I can't seem to find any reason why

[hallow flint]

in term of how effective a particular AD is

sounds like annotated data to me

[misty wharf]

@ocean wren but yeah definitely let me know :) thanks

[ocean wren]

She was originally going to do a templating project for me.. but she said she wanted to do some more ML work

Sounds like a nice project for her.. if she's not already figured out.. usually these students don't have a ton of imagination 🙂

I'm teaching houdini come october 🙂 should be fun

[misty wharf]

Yeah if something like this interests her we'd definitely be interested, we have a lot of demand for features that make it faster to build creatives and that's really the foundation of the whole tooling

[hallow flint]

are your lectures on youtube?

[misty wharf]

plus we have some big clients like NBC and Conde Nast so potential to see your work end up in use with some cool brands if she's into that sorta thing

[ocean wren]

no, we have to host them on the Uni video system.. which is terrible

I've got a meeting with her next tuesday, so I'll bring it up..

As with all student projects, hope for the best, but expect disaster 🙂

[misty wharf]

hehe, we've had interns from university of british columbia and they've all been pretty sharp :)

[ocean wren]

Yeah, its hit and miss to be honest

[misty wharf]

but even if it doesn't end up with anything production ready, even having a better idea on how it might be achievable would be useful

[ocean wren]

I've got one second year undergrad that went off to italy to do game AI for some company 🙂

[misty wharf]

right now we don't really have anything else than a vague idea of "this would be neat"

[ocean wren]

Yeah, it sounds like a neat use case for ML

[misty wharf]

You can ask her to email me at [e-mail] if she's interested

[ocean wren]

I believe she has a background in web UI, which is why it might entice her..

thanks.. I'll let you know if she bites with the idea.. honestly I don't like giving students contact until I've sussed them out. I'm supervising her over the summer, so will have a few weeks meetings before I inflict her on you 🙂

[misty wharf]

Alright :)

Feel free to mail me if you have questions about it yourself, can make sure I give you as official of an answer as possible lol

[ocean wren]

Yeah, no worries, much appreciated

And it sounds like a fun project

I've got a few of them running over summer.. NeRF vs Photogrammetry, various ML ones etc..

[misty wharf]

Nice

A friend of mine was using a drone to do photogrammetry, scanned an entire small building with it, looked pretty good

[ocean wren]

I rather stupidly took on a really low paying job as a loss leader.. which I now regret 🙂

Yeah, drone based photogrammetry and NeRF'ing are in my future this summer too

problem is, I can't run drones anywhere near here because we're surrounded by airfields

the UK's royal air force flies a ton of stuff in and around here

[misty wharf]

Oh, can't even fly at lower altitudes?

[ocean wren]

so no fun drone usage unless you drive for a while

nah, there's a 5km exclusion zone around airfields, and I think a bigger one around military airfields

and we have like 5 airfields in a ring around us 🙂

[misty wharf]

Ah

Yeah I think they have some rules about it here as well but not sure

[ocean wren]

This is where UK bomber command was based in WW2 for instance

[opal crest]

US is similar. Drone operation without a commercial license is extremely constrained (and the fines for violation are 5 figures...).

[ocean wren]

most of the UK drone force flies out of waddington, which is 5 minutes from my house

[celest python]

I also live very close to military airport in Turkey, no one cares when I fly a drone though

But still if they catch me I can get in a trouble 😄

[ocean wren]

yeah, its the catching and trouble bit I want to avoid

on the upside, we have a load of local firms who do anti-drone technology development 🙂

[celest python]

I'm just a regular 20 y.o. guy on their eyes but guess an university teacher shouldnt get caught like that 😅

[ocean wren]

yeah, it'd be bad 🙂

although, my colleage did capture a drone video of our building 🙂

https://www.youtube.com/watch?v=2ryVXPLM_Ds

[celest python]

can't you get a permission for a "research" reason?

[ocean wren]

not near military stuff

looks like he captured it late or early 🙂

[celest python]

Looks like a charm place

[ocean wren]

big curved glass building is where my office is

[celest python]

the place where has yellow lights?

[misty wharf]

uni here has a weird big ball outside of it lol

[ocean wren]

yeah, I don't think they're really yellow.. some weird optical thing on the glass maybe?

[opal crest]

Honestly don't know. StopMovement will stop the current move. If you get any sort of movement input after the stop movement, you might get behaviour like this. If you're really stuck, putting a breakpoint in your movement controller to see where the the movement is coming from and working your way backward could help.

[atomic coral]

hmm. it's really strange, it's not going anywhere that any movetos are telling it to, it's just slowly lurching forward

[visual silo]

How can I find a distance between two points on the nav mesh from blueprint? I expect I need these, but how do I get the relevant Nav Data? (And is there a more all-in-one node?)

[atomic coral]

are you asking how to find the points, or how to get the distance?

thinking my movement issue is somehow animation related

something in the abp

[runic flare]

I have a simple EQS query that generates a grid of points and selects a navigable point, setting the value of a vector blackboard key to that value

I have some really weird behaviour though -- the query keeps failing even though ALL these points are perfectly navigable

that on its own might not sound very strange, eg maybe it's my query logic etc

but the confusing part is that when I debug the EQS engine code, the query works fine and my agent can path towards the found point

the act of breaking and stepping through some lines of UEnvQueryTest_Pathfinding::RunTest makes the query work

[atomic coral]

that somehow sounds time related

[runic flare]

yeah strikes me as a race condition for sure

the issue obviously is that I can't tell why the query is failing because when I try to step through to debug the failure, it doesn't fail 😭

Im wondering if anyone has seen an issue like this before with the EQS, or maybe someone has ideas about what could cause a simple pathing test to fail

[atomic coral]

unfortunately, i've been meaning to try to figure out how to use EQS, but I haven't yet. so i'm not too much more help. is it possibly running the query before the navmesh is available? kinda spitballing.
does it work if you run at slomo 0.1 or some such?

[runic flare]

the navmesh is static and present -- after one EQS step-through subsequent queries stop working again, until I hit the breakpoint and step again

but that's a good point I'll try rebuilding the mesh just in case it's corrupt somehow

[atomic coral]

can you breakpoint the failure case, and work backwards from there?

[runic flare]

no, because when I break, it doesn't fail 😂

so Im not even sure what the failure case is

[ocean wren]

check the output logs?

add some printfs to the code?

[atomic coral]

so there's not a point in the code you're trying to debug through where there's like "return FAILURE" or something?

[runic flare]

yeah I'll have to add some prints

[ocean wren]

usually I instrument the code from input to output and intermediate steps and just look for patterns of failure

[runic flare]

this is all Im getting rn

[visual silo]

@sudden citrus those videos you posted from Bobby have been great. I've had to watch them like three times but I'm feeling really pleased with how easy this ai system is to work with and expand now that I've got the groundwork in place

10/10 would recommend

Would be nice to package it up into a marketplace plugin after I get a bit further..

[sudden citrus]

Cool! That's good progress

It's pretty nice to work with stuff when it's all separated out like that

though it did hurt my feelings when he said there's basically no gain to using an HTN... really wanted to try it out on an actual project

[visual silo]

It's interesting. This system seems to be a lot easier for both very complicated and very simple behaviors. Like it was incredibly simple to add a look-at react behavior; and then on the other hand I was able to make a much more complex investigate behavior that I was having trouble with in my previous HFSM / BT approach

and although I haven't implemented everything yet the way Bobby described regarding dynamic behavior modification, I've got a clear path for that.. and once that's in place it will be extremely simple to do things like have a statue that initially is asleep but will wake up and start chasing you if you're too noisy

[atomic coral]

yep looks like it's root motion... the animbp is playing a "stop" animation when the guy is supposed to come to rest, but that anim has some root motion, so that causes velocity, which causes it to start moving, then it stops, which causes it to play the stop animation, which causes velocity, which causes ... etc

[visual silo]

My major two critiques of the video is that:

1. Not everything is explained in great detail of what the benefit is to the design choices for authoring.
2. There's a lot of design choices made for efficiency-sake that conflate the issue for early prototyping

Both are actually the same issue. It's hard to know which choices are being made for workflow / vs optimizations.

But working through implementing it myself a lot of that is becoming clear

Oh, this reminds me. Does anyone have a good resource for Stealth AI audio barks?

[sudden citrus]

The idea is that the detail will come from your game's needs

[visual silo]

Sure, but there's already a lot of detail that comes from Bobby's experience shipping the system

[sudden citrus]

also IMO it's great for prototyping - I'm going to implement it at work fairly early on so that designers can get used to it and not paint themselves into a corner too quickly

[visual silo]

Yeah, 100% it is.

It's way easier to work with

Once the system is built, it's great

[sudden citrus]

as a digression this is also I hate the idea of just leaving 5 designers to it in blueprint only, because "it's quicker to iterate on" - no it's not, not after 3 weeks when you have a bug

[visual silo]

The way I built this up it's actually all still blueprint, but built on top of a AgentAIComponent

but every behavior, desire generator, and desire is a blueprint object

and the "actuation layer" is being handled by a BT

I think it could be strengthened for the kind of environment you're describing if the access pattern was more tightly controlled as Bobby describes

But would still have the ease of iteration of blueprint & BT

[sudden citrus]

I'd say use BTs as behaviours

the actuation layer is BT tasks

BTs sequence behaviours

[visual silo]

I suppose that's how I have it. There is a blueprint side to behaviors, but the purpose of that is to fill the blackboard with the needed info for the BT from the relevant goals

But the behavior decision making is outside BT

The BT itself is just a selector for ActiveBehavior == Patrol etc

Not really much other logic in there

and the blueprint side of the behavior handles target selection when for example you have multiple locations to investigate

I guess that last detail is why I'm thinking of BT as just the actuation. It's just the action-sequencer but is pretty dumb otherwise

The BTs role is basically as minimal as it can be

[sudden citrus]

Oh wait no

you don't want a single behavior tree for it

every behavior should be its own BT which you just run from the decision maker

and prefill the blackboard if you want

a "hardcoded selector" is a bad idea

the goal here is to let people script their own behaviours in a single BT per each, and not expose any assets that should be set in stone but are actually editable

[visual silo]

Right. The only reason I've got it all in one right now is because it gives me the ability to cancel a behavior midstream without me needing to basically rewrite that logic when switching BTs

Is it easy to do that? To stop a BT and start up a new one?

Having them actually be separate BTs would be better

[main mantle]

Hi, in UE5 I Have an NPC go to a truck to pick up an object and bring it back, it seems to have a prefered side and instead of taking the blue path it takes the red path (some times, it changes when I re-run sometimes). How can I make it pick the side of the truck closest to the NPCs start position?

[misty wharf]

you would probably need to make the acceptable radius take into account the truck's size or find the nearest point on it

if you just move to the truck's position it would most likely attempt to move "inside" the truck which may result in it pathing in an odd way

but this is just a guess, you can find out if that's indeed the case for example by drawing a debug sphere to the point where it's trying to move to and seeing if it's indeed inside the truck

[ocean wren]

ok, add another hate to the list.. I hate linux now too 🙂

[main mantle]

yes, I already have it pick a location radius of 100, without that it bumped into the truck and never got there. Ill do the debug sphere and see what it thinks

[ocean wren]

goddam stupid goddam OS

[main mantle]

@misty wharfThanks, I see the problem now 😄

[misty wharf]

was it moving somewhere weird?

[main mantle]

each time I pick a random spot in a radius from the center of the truck, some times that spot is on the other side of the center, some times on the closest side, now that I see it it is pretty obvious why 😄

[misty wharf]

heh :)

[main mantle]

and some times it does weird as stuff like this, the white sphere si the radius of the random location, the black vertical lines is the random spot it picked, the start pos is the red blob, but about 10 times as distant, and the NPC has gone to the front of the truck 🤣

But it's cool, I can fix it now

[final trail]

is it possible to tell ai sight perception to ignore certain components?

[misty wharf]

Kinda depends on what you mean by ignore

You can implement IAISightTargetInterface on your actors to customize how the sight check is done on a per actor type basis

You can also just do checks in the perception updates

[final trail]

my character has a sphere attached to it, and i think the ai is detecting it

[ocean wren]

you probably need to change the sphere's physics to not accept visibility traces

[spring inlet]

you could add some sockets to the mesh of the truck, e.g. on the left/right door, find that socket location and use that as target location

[main mantle]

I have done that with another actor and it works quite well

[glacial vale]

Hello guys, I'm trying to work with "Dynamic behavior Trees".... I found that I have to use behavior tree component (probably on the AI controller) my problem is that I no longer can set blackboard value from my character or controller... Any idea how I should do that?

[ocean wren]

finally got pytorch setup on my A6000 dev box

and have been generating images all afternoon (as you do)

Disco diffusion prompts are different than others, has a sort of painterly feel to it

My usual "man walking to brutalist megastructures" kind of thing 🙂

Diffusion models are so cool

Reminds me a bit of a sort of high-res guild wars 2 vibe 🙂

[celest python]

it's all abandoned places

less humans, more environment

and always combination of a specific types, i guess?

algorithm shows itself

[hallow flint]

man these are so cool, was this dev machine being setup in linux? @ocean wren

how much time are these taking to generate?

[sudden citrus]

Yeah you just run the new one as you would on begin play

[visual silo]

And that aborts behaviors on the current running BT?

[sudden citrus]

Yes

You can test it pretty easily

[ocean wren]

That's just the prompts I gave the generator.. I avoid humans because its not great at them (other models are way better)

[ocean wren]

Linux yeah, although this is a colab notebook so would run in the cloud too (I run the colab locally on my dev box with an A6000 in it

[hallow flint]

you using ubuntu?

[ocean wren]

Generation time is a tricky one, because its a diffusion model, so it refines over time. This particular model does take a while to get any detail, but probably about a couple minutes currently?

yeah, ubuntu

[hallow flint]

okay that's not too terrible, i need to give this model a try, those results are impressive

[ocean wren]

if you look for "disco diffusion" you'll find a colab notebook, you can run it as-is

I just wanted to experiment a bit and not run out of colab notebook time 🙂

Going to try some 1920x1080 generation today

it also does animation, but I've not tried that yet

[hallow flint]

Yeah i looked it up earlier. I need to get on my machine with GPUs before i try and do that

animations?

😮

so you could technically get a few frames of those cool environments you made

[ocean wren]

yeah, you basically give it a camera path

there's a bunch of tutorials on youtube

[hallow flint]

checking for one right now

[ocean wren]

I mean a load of them look like crap 🙂 but you CAN get excellent results

[hallow flint]

haha

well

let me know when you get that pipeline working of going from generated concepts to actual 3d worlds

[ocean wren]

I'll be doing a demo video for the funded project where we show some of the ideas in prototype

[hallow flint]

i assume that probably won't be released publicly right

[ocean wren]

will do it in our new mocap room 🙂 should be fun

well, its a prototype, so probably not.. but the company I work with gets a 12 month option on exploiting it

[hallow flint]

ahh

that's pretty awesome

[ocean wren]

https://www.youtube.com/watch?v=kexc6bA0AZ8

[hallow flint]

this is so doom-like

awesome

[ocean wren]

ooh, this guy has more

https://www.youtube.com/watch?v=Qda6nuhQxcA

[hallow flint]

i saw a music video generated with one of these models, i dont think it was disco diffusion though

[ocean wren]

Yeah, there's been a bunch of those recently

I saw a really nice one yesterday

ooh yeah, this guy has some really nice ones 🙂

[hallow flint]

i found it, it was also made using disco diffusion, what are the odds lol

[ocean wren]

Disco has the advantage that it doesn't just do square images

I'm curious what resolutions it can get too.. will stretch my 48gb card today 🙂

[hallow flint]

hahaha

well let us know what kind of results you end up getting

[ocean wren]

probably terrible 🙂

[hallow flint]

all part of the process haha

[ocean wren]

Good prompt tuto here: https://www.youtube.com/watch?v=w7VFkU1zM30

I like the painterly look of this model.. obviously trained on a lot of artstationy stuff 🙂

[hallow flint]

Remember how you were telling me that you could use a GNN for setting the relationships between objects in an image and could directly use that as a rule-based placement module in videogames

is that already being done?

[ocean wren]

no, thats the current work, gnn's are generally being used though

[hallow flint]

im saving this, thanks!

[ocean wren]

GNN's are good for things where you want to prompt for completion, for instance they're used to suggest similar images on pinterest

and google uses them to suggest routes in google maps

[hallow flint]

we've used them for some network problems

[ocean wren]

they learn the structure of content and can then give probabilities for next most likely values

So the innovation is to just learn via an ensemble a scene description in the abstract and learn over that using the GNN

[hallow flint]

so just to further clarify that for myself, the output of the gnn will be the edge relationships, corret?

edge-level classification?

[ocean wren]

it'll be a graph recommendation

basically a scene description in another graph form

an extended PROSE language description

prose being a cinematography language

[hallow flint]

ahh okay, i was googling it but i couldn't get anything results that made sense lol

[ocean wren]

yeah, there's no current approach doing this

just parts

the closest is a paper called MetaSim2 by someone in Sanja Fiedlers lab at U Toronto

where they generate an unreal engine scene from a grammar

[hallow flint]

ahhh interesting

[ocean wren]

for creating driving simulation training data

[hallow flint]

im having some trouble understanding certain concepts here

[ocean wren]

they used RL, but I'm taking a different approach to the graph generation part

[hallow flint]

so your output would be somewhat similar to this?

[ocean wren]

its basically a graph probability "given graph X, what would the resultant complete graph X' be"

[hallow flint]

i found the paper

[ocean wren]

yeah, only for metahumans and scene content in general

[hallow flint]

Okay so i guess the part that im a little confused is

what would be the input to the GNNs

so you'd give it a description of the components you extract for your ensemble of models, those would be your nodes

[ocean wren]

the input to the GNN's is a timeline of graphs which describe each frame of a film

[hallow flint]

I assume you wouldn't provide edge information?

[ocean wren]

yeah, you would, the edges are the relations between the entities

so for instance, you have a scene in a movie where two actors are arguing

the graph would have actor A, actor B and edge relation "arguing"

[hallow flint]

right

[ocean wren]

now obviously, there's more in a scene.. but you get the picture.. might have object "glass of water" with relation "on top of" and object "desk"

[hallow flint]

yup

and all those would be provided to the gnn, that i get

[ocean wren]

So we build as complete a graph as possible, for every frame

forms a timeline of graphs, basically just a whole load of spatio-temporal graphs

and then feed that to the GNN

[hallow flint]

yup i follow so far

[ocean wren]

then at runtime, you provide a partial graph in the editor.. say you add a desk to the scene

the partial graph prompts the GNN for completion and it has high probability of "glass of water" with relation "on top of" and object "desk" so it suggests a glass of water should be on the desk

[hallow flint]

ahhhhhhhhhhh

i understand

[ocean wren]

So as you edit the scene, it reworks the probabilities for graph expansion

based on the scenes it was trained on

[hallow flint]

I see

that's super neat

and you could directly grab that output and use it for automation of asset placement

[ocean wren]

The biggest hurdle, is that there's lots of noise in the initial graphs, because the ML models aren't 100% reliable yet

[hallow flint]

i know we talked about datasets for these kind of problems before and how there exist frame by frame annotation

[ocean wren]

yeah, well, in this case we use the GNN to generate a graph in a new grammar called Scene Description Language, which is an extended PROSE grammar (prose came from Rene Ronfard)

[hallow flint]

does that also include the metadata necessary for relationship between entities in the frames?

[ocean wren]

and then parse the SDL into an unreal engine generator and bishbashbosh

yeah, it includes the relationships and how they change over time

[hallow flint]

ohh i see

[ocean wren]

usually they wont change

[hallow flint]

that's so cool

[ocean wren]

glass on top of desk is unlikely to change, unless someone picks up the glass and throws it 🙂

[hallow flint]

can we expect this to be open source in 2024?

lol

[ocean wren]

So the funky part, is that all of this can be done using natural language

thanks to huggingface and others

[hallow flint]

i love hugginface

[hallow flint]

explain this to me?

[ocean wren]

in a rather convoluted pipeline that goes something like this...

you take your voice from the microphone

pass it to azure cloud to do speech to text

pass the text to huggingface to do entity extraction etc

use that as the input prompt to the GNN

and then the GNN generates the SDL and the SDL gets parsed and rendered

the huggingface stuff might also be microsoft's gpt3 language model too.. havent fixed that one yet

[hallow flint]

is this really as effective as it sounds?

[ocean wren]

of course not 🙂

[hallow flint]

haha

[ocean wren]

its like.. you ever seen steamboat willy? the really early disney animation?

[hallow flint]

yeah i have

lol

[ocean wren]

yeah, so you know its utter crap right?

[hallow flint]

so

yup lol

here's what im wondering

[ocean wren]

well, we're at that point with this stuff... its going to be janky and not work a lot

but will improve rapidly

[hallow flint]

I wonder if you could describe a scene in whatever way you want. And then use gpt-3 to parse it into SDL format

[hallow flint]

yeah it's pretty awesome seeing how all this stuff changes so rapidly

[ocean wren]

I guess if you could fine-tune gpt3 on SDL examples.. but gpt3 takes a supercomputer to host 🙂

[hallow flint]

Yeah that's true

[ocean wren]

there are language models being trained for open source.. currently a 175bn param one

[hallow flint]

but who knows, maybe gpt3 is really good at one particular sdl type where you could get that sort of description

[ocean wren]

I reckon fine-tuning could work.. but thats a project for later

honestly, just getting a few examples working is the main goal

[hallow flint]

you're talking about umm gpt-j or one of its variants?

[ocean wren]

enough to give a demo and record it 🙂

[hallow flint]

eleutherAI i think is the group

[ocean wren]

no, there's a uni with a big cluster doing it.. can't remember the name of the project

but they're doing it specifically for open source access

[hallow flint]

not this one? https://www.eleuther.ai/

[ocean wren]

so no company

no, thats a different one

there's a few 🙂

[hallow flint]

these guys are also doing open source ones

[ocean wren]

eleuther are doing image generators too 🙂

[hallow flint]

yeah that'd be super cool

it's so hard to keep track of this

you use something to track latest and greatest research?

i've been told to just follow certain people on twitter

[ocean wren]

semantic scholar, twitter 🙂

[hallow flint]

but haven't gotten to do that yet

haha

[ocean wren]

paperswithcode.com

that kind of thing

[hallow flint]

ah got ya

[ocean wren]

I usually have a pass over papers from the main conferences like CVPR, NeurIPS and Siggraph

but usually I let others do that and I follow them on twitter or youtube or whatever

[hallow flint]

i go through the titles of those presentations as well

[ocean wren]

There's a few key academics whose research groups all post stuff on youtube, so I follow their labs

people like Michael Black, Andreas Gieger, Matteas Neissner, Sanja Fidler etc.. basically profs that I like 🙂

[hallow flint]

for PCG?

[ocean wren]

kind of a mix of things.. Michael Black's lab does PCG for humans

Gieger and Neissner are both ML + geometry guys

Fidler is working with NVIDIA on interesting ML stuff

I also attend conferences virtually

like FDG and Siggraph and the like

[hallow flint]

i need to do mroe of that

siggraph is for members only, right?

or are those free

[ocean wren]

you can usually get free access to a tiny bit of the content

but yeah, its a huge graphics conf, so worth paying for

[hallow flint]

does your university coffer fees if you want to get the full experience?

or access to content thats paid?

[ocean wren]

used to.. but not so much now

have to pay for myself, or get some grant money from somewhere

[hallow flint]

[ocean wren]

if you've got a paper at a conf, they'd usually pay

[hallow flint]

i am going to ask for money for memberships

[ocean wren]

BUT, you have to pay the goddam conf/journal to publish.. its so crazy

[hallow flint]

really??

[ocean wren]

So they charge you money, for you to do the work..

[hallow flint]

wth

[ocean wren]

they say its for "publishing" costs

and admin

which is horseshit

because we all review for these things for free

and everything is electronic anyway

its how elsevier and springer make so much profit.. total scam

[hallow flint]

and then some publishers hide the papers behind paywalls

[ocean wren]

exactly

[hallow flint]

and i've been told that the author doesn't even receive anything if someone ends up paying for it

insane

[ocean wren]

although, to be fair, most are publically available, because these days funders require versions of papers to be public access

yeah, you don't get diddly

its completely crazy

But then the politics of academia are crazy anyway

[hallow flint]

you've talked me out of ever wanting to get into academia

lol

[ocean wren]

Save yourself 🙂

[hallow flint]

you have a phD right?

[ocean wren]

No, this is me completing it 🙂

[hallow flint]

ohhh

[ocean wren]

came from the games industry

[hallow flint]

how long do you have left?

[ocean wren]

a couple years

[hallow flint]

so excuse my ignorance here so correct me if im asking the wrong question

do phD's need a thesis? and if so, what is yours?

[ocean wren]

yeah, there's different types of PhD, by publication (a summary of a bunch of papers) or by thesis (a big thesis)

My thesis title is long and I can never remember it, but it is basically about generating emotional digital actors 🙂

[hallow flint]

hahaha

emotional digital actors

nice

[granite vault]

Interesting

[ocean wren]

basically actors for the holodeck

[granite vault]

What's that suppose to do with holodeck?

[ocean wren]

i.e. janet murray's version of a new immersive media

[hallow flint]

speaking of the holodeck

i remember talking to some defense people

and they were basically asking for someoen to make the holodeck

lol

[ocean wren]

holodeck is shorthand for immersive virtual world 🙂

[granite vault]

Google says the Startrek TV series thing

[ocean wren]

I don't mean actually holographic rendering or anything

[hallow flint]

how many years into your thesis do you have so far?

[ocean wren]

yeah, think of it like the star trek thing, a world where you have an utterly believable surrounding, but where that could be anything

good question.. I've been working on this stuff since 2004 🙂

but technically, I'm 2 years in

[hallow flint]

wow, that's a long time. I assume in your research you must have tried a few different ideas/approaches for emotional humans right?

[ocean wren]

I had a long time to think about it, worked on some adjacent projects, but started the PhD at my new uni two years back

[hallow flint]

ahh okay

So when you say emotional humans

[ocean wren]

Yeah, I was trying to use the traditional game AI methods at first

i.e. create an emotional model and do some game ai characters that can enact them via animation

[hallow flint]

are we referring to demonstration emotional expressions through animations, speech, text or all of the above?

[ocean wren]

I worked on Ground Branch for a bit to understand how people respond to movement

but ultimately, I came to the conclusion that game AI approach is too hard and ML is likely to be more of a factor

all of the above yeah

multi-model display

[hallow flint]

ohhh nice

[ocean wren]

I mean think of a less stupid looking version of Michael Mateas's Facade 🙂

that was my initial approach

[hallow flint]

I'd like to hear more about this if you don't mind

what was your emotional model compose off, did you have a list of triggers to emotions (things happening in the environment, that could trigger specific emotions?)

[hallow flint]

haha i have a friend who is very interested in facade

[ocean wren]

https://www.youtube.com/watch?v=POv1cOX8xUM

I mean it has some interesting ideas, but I can't deal with the graphics 🙂

[hallow flint]

i've seen this video a few times haha

[ocean wren]

its probably still state of the art sadly

[hallow flint]

well with gpt-3 being out

i've seen some pretty cool stuff being done with that

giving game ai a little bit of personality purely through seeding the ai with an initial prompt

"you are construction worker having a bad day..."

and that sparking off personality when you try and interact with that AI

[ocean wren]

yeah, there's some good snippets of useful stuff out there

but fundamentally, its just too much work 🙂

[hallow flint]

oh yeah

[ocean wren]

https://arxiv.org/pdf/2009.00541.pdf

Microsoft paper on why 🙂

"Its unweildy and it takes a lot of time" 🙂 perfect assessment imho

[granite vault]

Don't mean to interupt the conversation. Giving Ai personality are usually including dynamic dialogue isn't it?

[ocean wren]

yeah, there's lots of things you need

[granite vault]

And i think there was a chat about it here on having million dialogues on Valve's old game?

How is it even possible to store million dialogues?

[ocean wren]

yeah, valve's dialogue system was good, it combined voice lines and animations along with reactive prompts

[granite vault]

Can UE even store millions of dialogues though?

[ocean wren]

millions of dialogues, try billions 🙂

[hallow flint]

this is a good read

[ocean wren]

modern ML language models work across billions of items

[granite vault]

Hmm..

Forget about million dialogues. I still unsure whether UE can store million vectors with less memory.

[ocean wren]

if you look at large language models like GPT3

part of the value of ML, is that it is reasonably good at dimensionality reduction

you learn whats called a latent space, which is fewer dimensions than the original inputs

and it learns to cluster similar items

so you get a representation that's pretty small, but represents a huge amount of data

[granite vault]

Hmm.. too advance for my brain

[ocean wren]

Here's an eeasy example.. imagine we have a sentence "the cat sat on the mat", we could turn each word into an index value 0,1,2,3,4 etc

[hallow flint]

oh speaking of models and games, what does your current model <-> unreal pipeline look like @ocean wren ?

[granite vault]

Wait. If that really works well. Then even much easier to store billions Vector3 in UE with less memory?

[ocean wren]

each "the" would have the same index value

[ocean wren]

You wouldn't store vectors, you'd store a model in ML form and generate the vectors

[granite vault]

I think i get it that part. As theory. Thanks

[ocean wren]

it wouldn't always generate the exact thing, but it'd stastically generate something similar

[granite vault]

What is it like for UE's version of ML form, please?

[ocean wren]

UE doesn't really have it yet

[granite vault]

Uh too bad then

Thank you. I won't interupt your chat with Josue any further. My apologies

[hallow flint]

haha no worries

[ocean wren]

Its complicated.. its a few thousand movies -> into a kubernetes cluster running dozens of ML models over each frame, into a OpenTimeLineIO timeline of graphs

[ocean wren]

This is stuff I'm working on currently

[hallow flint]

Well there is that Neural Network Inference (NNI) plugin, that allows you to load models into unreal, but there's still work you have to do

[hallow flint]

did you build out the kubernets cluster training framework?

the reason i ask

[ocean wren]

Yeah, until its properly debugged and documented, I wouldn't bother 🙂

[hallow flint]

in this paper you linked

some RL reserach mentioned they started out with a kubernets framework and then switched over to rllib

[granite vault]

Ping me once you launch it or something. Interested to see though i am not sure what i could use for.

[hallow flint]

i don't know much about rllib

[hallow flint]

i already spent like 3 hours going through the source code...

[ocean wren]

Yeah, I am still finishing it, I mostly kubernetesized the MovieNet codebase and extended it

[hallow flint]

hahahaha

im used to working with code with 0 documentation

i work in a small company where people don't like to write documentation lmao

[ocean wren]

Hahaha.. but that's just bad for an engine company

[hallow flint]

i know

[ocean wren]

plus.. my bet is that it won't actually work most of the time 🙂

Unity ML agents has an issue with the same thing

kind of works, kind of doesn't

[hallow flint]

which part won't work?

are you talking about the parallelization of training?

[ocean wren]

I'm not doing RL though, for this at least.. so its easier to kubeflow it 🙂

[granite vault]

Wait. Marketplace is selling some NN plugins to let you load such things i just recall

[hallow flint]

yeah they have one that is simillar to the one that ue5 just released, that one is documented

i think it was a bit pricey though

[granite vault]

But like Josue said, need some work still

[ocean wren]

Had a student working on demonstration learning for level playing.. some of it worked, some of it seems broken

[granite vault]

Uh i see

[hallow flint]

ohhh got ya

i'm interesting in parallelization for RL training for multiagents

[ocean wren]

I don't think its a great idea to run ML models on the same GPU you're rendering with 🙂

[granite vault]

Did you buy it btw? @hallow flint

[ocean wren]

its probably fine at edit time, but I wouldn't do it in a game yet

[hallow flint]

i did not, i just looked at the ue5 released version

[granite vault]

Noted.

[ocean wren]

The Ark Raiders guys wrote their own kubernetes RL training stuff

[granite vault]

You pinged wrongly

[hallow flint]

yeah i think you can run headless though. But I haven't actually done any metrics

[granite vault]

Yeah. I saw Embark studios, right? Bunch of EA former staffs too

[hallow flint]

im building one too

but now after going through this paper

[ocean wren]

yeah, the SEED EA R&D team got poached 🙂

[hallow flint]

im wondering if i should check out this rllib stuff

[ocean wren]

check it out and report back 🙂

[granite vault]

One of them was in this UE discord btw. I think he left this discord because he is seeing too much crazy questions like making game into gpu which should not be.

[hallow flint]

I will, ill read up on it tomorrow

[granite vault]

He is really genius. He made that volumetric cloud in old BattleField games.

[hallow flint]

speaking of building training framework and pipelines, i was reading this gdc talk

might be of interest

https://www.gdcvault.com/play/1027631/Machine-Learning-Summit-It-s#

The tldr; is

well let me grab a relevant screenshot

for those that are interested in architecture

[granite vault]

How does that work? They store all the training into controller?

Ubisoft show cased the AI animation with nice physic

[hallow flint]

In the first image. The training and storing of the model is happening outside the game Engine.

[ocean wren]

I don't see what advantage that brings?

[hallow flint]

On the second image, they basically said that it's better if that model and the underlying framework should happen inside the engine

[ocean wren]

other than eating GPU 🙂

[hallow flint]

I believe the advantage of the second architecture is that it facilitates things in production?

i'd have to give it a second read

[ocean wren]

Hmm, but you'd have to copy your weights as they train

so you'd have a copy on both ends

[hallow flint]

Right, i think this is a way of abstracting the training/metrics away from the gameplay logic

So I'm personally doing it the way the first picture is doing it

[ocean wren]

and to make a model trained in python into a runtime thing, you just run it via torch C++ 🙂

yeah, the first method seems so much more rational

its also how Unity agents does it

[hallow flint]

well now you can just run it in onnx too, all built into unreal

[hallow flint]

Yeah i think so too. well im not building agents for videogames, i hope to eventually try them out on real robots

[ocean wren]

true, same idea though right.. build a trained model, load it from some optimized version and run it via C++

[hallow flint]

yeah

[ocean wren]

if you want to train robots, look at MuJoCo

recently open sourced by deepmind

[hallow flint]

Oh we don't really focus on low level control

we mostly do high level behavior training

high level tasking

[ocean wren]

gotcha

[glacial vale]

I thought I had a simple question behavior tree related or maybe I didn't formulate it properly but... I'd love some help

#gameplay-ai message

TLDR: If I use behavior tree component on my AI controller how can I use "Set Blackboard Value" from by pawn or AI.... It seems I can't use the "get blackboard" node anymore ?

[ocean wren]

Use "GetAIController" on the character and then call a function on it to set the BB value

SetBlackboardValue only works on AIController because its got a blackboard, the other classes like character and pawn don't

so you need to get the AIController that is controlling them and ask it to set the BB value

[glacial vale]

this work from a pawn for example

[ocean wren]

So what is the difference in types of actor from your first example to your second?

You can't just stick a brain component on anything

the framework has some C++ code to handle the component creation

its a bit of a weird design imho

[glacial vale]

The easy way is to use behavior tree like this it works no problem.... I have a hard time using the behavior tree component... I have no problem digging in the C++ but I thought it was not required here

[ocean wren]

Well, if you use the usual setup of Character + AIController than it'll work

or Pawn + AIController

anything else, you'll need to dig into the code to see how its setup in C++

[glacial vale]

I have a character + AI

my amazing implementation in AI Controller 🙂

it's just I can't seem to be able to set any blackboard values

[ocean wren]

What error do you get?

[glacial vale]

Blueprint Runtime Error: "Accessed None trying to read property Blackboard". Node: Set Value as Object Graph: EventGraph Function: Execute Ubergraph BP AI Controller Blueprint: BP_AI_Controller

[ocean wren]

Where does start logic come from?

why not run behavior tree?

[ocean wren]

this example should definitely work

why not "get blackboard" node?

[glacial vale]

Behavior tree component is the only way to work with dynamic subtrees

[ocean wren]

Hmm, I haven't played with that myself

[glacial vale]

I'm guessing that "run behavior tree" kinda add a behavior tree component ...

not sure

[hallow flint]

@ocean wren okay so now i see the benefits of using rllib for setting up your training framework. it already provides a custom template for connecting to your external (unity, unreal, or other game engine) directly into their extensible framework. This is pretty neat. It handles a lot of the troublesome issues with networking that are usually involved. https://docs.ray.io/en/latest/rllib/rllib-env.html#external-agents-and-applications

[ocean wren]

Yeah, apparently it runs in kubernetes too

But not really that useful for my purposes to be fair, I'll pass it onto my students who express an interest though 😉

[hallow flint]

haha i might play with this on my own free time

[ocean wren]

my own use case is far more about video throughput and ML efficiency than it is the networking bit 🙂

[hallow flint]

i am already too committed to my other solution haha

ahhh

i have been fighting with docker networking so much this last week

not fun

[ocean wren]

plus, I don't trust code until its been battle tested for a while

yeah, docker is a pain

k3s, k3d ugh 🙂

[hallow flint]

i fix one networking thing in it, and another breaks

[ocean wren]

portainer

and rancher

[hallow flint]

i haven't tried this out, should i be using this?

[ocean wren]

had some frustration yesterday trying to get nvidia drivers and pytorch working properly

[hallow flint]

i was reading about portainer

sounds so convenient

[ocean wren]

k3s is basically a mini kubernetes env

[hallow flint]

i am getting tired of doing everything in the terminal

my brain can only hold so many linux commands

[ocean wren]

for low-power device work, but good enough for my purposes

[hallow flint]

oh god

that is the worse

i've literally broken a few computer at work

because of gpu drivers

i just threw them back to IT

[ocean wren]

its amazing, so many people doing linux ML using nvidia and yet its like the most janky shit in the universe to install 🙂

[hallow flint]

i know right

as someoen working in linux all the time

[ocean wren]

its almost like python "engineers" are unable to do ux 🙂

[hallow flint]

im used to spending a bunch of hours on simple shit

loll

[ocean wren]

I hate it with a passion

[hallow flint]

i have no choice

[ocean wren]

all of these stupid fucking arcane command line things

[hallow flint]

i've also been hating windows lately

[ocean wren]

lhsw -C network what???

[hallow flint]

some stuff i can do easily in linux i can't do in windows

and vice versa

[hallow flint]

yes

[ocean wren]

yeah, its the worst of both worlds

[hallow flint]

hahahaha

[ocean wren]

usability constrained by stupidity and functionality constrained by usability 🙂

[hallow flint]

do you use conda?

or .env

[ocean wren]

I do conda

I also pip

but I mostly conda

[hallow flint]

okay same here

[ocean wren]

I'd nuke it all from orbit if I could

[hallow flint]

🤣

[ocean wren]

along with python

jettisoned into the sun

[hallow flint]

https://tenor.com/view/bye-chat-goodbye-bye-burmese-python-python-gif-23678463

[ocean wren]

its all just bad

[hallow flint]

this is python saying goodbye

lol

[ocean wren]

why can't we have a NICE language

like C#

[hallow flint]

my manager really likes c#

[ocean wren]

or better yet, no language at all

just freestyle it 🙂

[hallow flint]

maybe we'll get to the point where we can just speak out our text

[ocean wren]

its coming

we won't have to program at all soon enough

[hallow flint]

maybe gitub copilot will get us there

i forgot the other one that is even better

[ocean wren]

yeah, there's a few of those going around

me too.. can't remember the name 🙂

[hallow flint]

openai codex

[ocean wren]

we tried it at work the other day

works well

[hallow flint]

"parse my images to fit this nn right here please and ty"

which?

[ocean wren]

dunno 🙂

someone else was demoing it

might have been a new one.. hush hush and all that for all I know 🙂

[hallow flint]

haha

[ocean wren]

point being, that these large language models are getting crazy powerful on many fronts and hopefully we see good come out of them and not just crazy capitalist evil for once 🙂

[hallow flint]

when i was googling i think i saw some articles saying that people will have to pay for them

[ocean wren]

Yeah, like most things, there'll be gatekeepers extracting rent

and running around in superyaghts 🙂

I spent yesterday cursing at videos of superyachts

you know, the type that cost half a billion dollars

and the owner has two or three

because that's the world we're in

[hallow flint]

lmao yeah it's insane

i think i saw a video recently of someoen asking these owners what they do

most of them had inherited their wealth

but one guy said he started out by cleaning boats

[ocean wren]

most of the current ones stole it from the russian people

[hallow flint]

and built his way up

[ocean wren]

yeah, I don't believe that 🙂

[hallow flint]

he could have totally been bsing

[ocean wren]

its probably a nice narrative he tells himself

[hallow flint]

loll

[ocean wren]

bootstraps and all that

I'm sure Bill Gates has the same narrative

just forgetting his family was fabulously wealthy and gave him a headstart

[hallow flint]

i've heard that same story about many people that 'made their wealth'

lol

[ocean wren]

Its the people like Abramovic that get me.. basically putins buddies who were given oil/gas companies and who acted as bagmen for putin

NOBODY imho, should have that kind of wealth

[hallow flint]

i don't really follow politics im not sure who abramovic is

[ocean wren]

he's a russian oligarch that was a pal of putin, had a football club here in the UK until recently

owned an oil company I think? something like that

[hallow flint]

wow

how did he lose his club, did they take it away from him

or he sold it off

[ocean wren]

technically he still owns it

but he doesn't control it

[hallow flint]

ahhh okay

[ocean wren]

he handed it over to a trust

[hallow flint]

did UK freeze his assets?

oh i see

[ocean wren]

its complicated 🙂

he got sanctioned if I remember rightly

but it really doesn't matter when you're that wealthy

[hallow flint]

yeah it's pennies to those people

[ocean wren]

because you can get on your jet and travel to one of your many superyachts

this guy has three? I think.. all like 500 million each

[hallow flint]

wtf

[ocean wren]

ah, yeah, his fleet cost 2 billion

https://www.youtube.com/watch?v=6kTR8BXvqNo

[hallow flint]

12.9 billion

holy shit

[ocean wren]

oil money stolen from russia

imagine how many gpu's he could buy 🙂

[hallow flint]

probably as much GPUs as OpenAI gym has

hahaha

i wonder wh ythey need so many yachts for

i think these have an insane cost just to maintain

[ocean wren]

yeah, if you watch that channel, you'll find out a load about it

I binged yesterday 🙂

[hallow flint]

haha

went into the youtube hole

[ocean wren]

well, the guy on the channel works on them

so you get some good insights

[hallow flint]

ah i didn't even realize the channel itself is called eSysman SuperYachsts

lol

[ocean wren]

I mean they're beautiful boats, but totally and utterly inconcievable to me

that anyone should ever amass enough wealth to own one

[hallow flint]

yeah im with you, not like they even live on them

[ocean wren]

Very rarely

most of them spend like 2 weeks a year on one

[hallow flint]

okay im going to bed, it's super late here. Good chat see ya

[ocean wren]

night 🙂

early doors here 🙂 time for a snooze 🙂

[noble hornet]

Heyo, quick question. Is it bad practice to combine multiple tasks into one. So for example, instead of having two separate tasks where a unit focuses a target and then moves to it. I'd combine it into one.

[ocean wren]

no easy answer there.. I was at an industry panel a good number of years back and we had multiple guys present theirs.. there was no concensus

some had like 5-6 different complex behaviours

others had hundreds of simpler ones

So, I guess the answer is.. what do you prefer?

this was at a game ai conf in vienna with most of the european companies (EA, Ubi, Crytek etc)

Mieszko from Epic was there 🙂

[celest python]

I guess I can adapt to anything but not Ubisoft's way 🤣

[final trail]

hi, im on ue5 and it doesnt let me select ai controller class, anyone know why?

[misty wharf]

Select it where specifically?

[final trail]

In the actor self tab

[misty wharf]

I'm not quite sure what that is, can you take a screenshot showing it

[final trail]

In a bit

[celest python]

so your class doesnt appear there?

[final trail]

it does

it doesnt allow me to select it

[celest python]

have you tried restarting the editor? its might be just editor bugged out

[final trail]

like as in closing and reopening

if so i already did that

[celest python]

can you select it if you create a new fresh class

[final trail]

nope

[misty wharf]

🤔

5.0.2?

[final trail]

correct

[celest python]

what happens when you select

it just resets to "none"?

[final trail]

it stays at none

but when i click on a built in class it works

[celest python]

try pressing X and select again

its an editor issue probably

[final trail]

didnt work

that's unfortunate

[celest python]

for a workaround you can set in on construction script

set ai controller class function

not sure if its exposed to blueprints though

[final trail]

That could work

I’ll try it when my power come back lol

[misty wharf]

I would try asking on #ue5-general, maybe there's more people who might've ran into this just not on this channel

Seems like a pretty major bug

[final trail]

the work around worked

[slim temple]

how can i get my ai to repeat the path patrols after reaching the last one?

[misty wharf]

Depends on how you have it set up, but you could for example set the next target back to the start point, or go through the points in reverse

[slim temple]

not sure if reverse will work for what ive used it for

[misty wharf]

Yeah I guess you could just try going back to the first point then

[slim temple]

how could i do that? 🙂

this is my BT

[misty wharf]

Probably update your increment path index task so it handles it correctly

[slim temple]

wdum

[misty wharf]

Well you could have it check if the index is at the final point and instead set it to 0

[slim temple]

where would i add that

sorry i am new to all of this

[misty wharf]

It looks like you're already checking if you're at the last point in there I think?