Procedural Islands of 'Dauntless'

[Music] so welcome to our talk thanks for coming it's a 5:30 which is late in the day so I appreciate it so first off my name is Michael trache I prior to this spent a bunch of years 15 that Bower on Knights of the Old Republic an artist was the Louisville art lead through massive X 1 through 3 and in the early days of what is now anthem I was a technical art director hi my name is Nicola Koenig and I'm an engineer at Phoenix loves and previous their work that side effects on Virginia and Houdini engine plugin for unreal so today we're gonna talk a little bit about Dauntless the game we're working on right now we're going to talk about what we wanted to accomplish where we're at and to those that don't know we're

discussing a Houdini Houdini engine pipeline so we need to cover a few basics on just some Houdini terms so that some of this makes sense to those with less experience with that and from there we're gonna describe our overall pipeline what we've set up it's a still work-in-progress what we're gonna describe where we're at and we're going to dig into a few sort of useful pieces for anyone interested in sort of a Houdini engine pipeline of their own so that's those building blocks so Dauntless is a co-op action RPG it's a free-to-play title that's currently in early access you'll fight powerful behemoths and craft weapons and play with your friends so it all takes place on the shattered Isles which is pretty much the topic of our talk here so if you want to learn more about dauntless as a game and everything else invite you to go check it out at play dauntless

calm but the rest of this should hopefully become self-explanatory so yeah the shattered Isles they play a pretty critical role they're not just a passive backdrop like some we joined the team because working on this a pretty exciting environment was was a lot of fun personally but also because we're doing this with a really really small team and that meant we couldn't go about this basically in the traditional way when I say small team most of the world art you're gonna be seeing here was created by world artists with Jen Morgan technical art with Cory Lake and our direction by Glenn Barnes so very few hands were involved in this so when I say instead of talking about an art department I'll probably say Jen or Cory and think of them as synonymous we wanted to we had an open-ended you know

list of themes biomes that we wanted to kind of take on we only had some concepts and some idea of the types of environment to take on when we started this but we knew that there'd be more to come and it'd be an evolving kind of thing an evolving digital target but also just you know we'd kind of adapt to new themes as they showed up so we didn't know the scope of everything that we'd be taking on when we started it but visually we knew there'd be pretty wide range technically we couldn't just make do with any space we we wanted fairly predictable topology for you know the player navigation gameplay but these these are procedural spaces and we'll get into some of what that means but they're also directed spaces that means the level craters I'll say designers but it was sometimes Jen but you know everyone sort of contributed to that but the idea is that we wanted to kind of

lay out what would happen and that would be part of the input so that would mean that we land on the island we start in this swamp when we make our way up into this forest area over here whatever that'd be the type of input that some of them might want to sort of begin creating some of these things with we haven't ruled out like generating just sort of new things that and see how they go but in general these are a directed thing we needed to be compatible with unreal we wanted Jen - Jen and kori to continue working on an island as if it was created by hand so the end result of this like specifically that means you know maybe not a lot of custom or custom mesh we had to be careful with because we wanted her to resume painting landscape or plants or manipulating meshes and that kind of thing and her her overall workflow couldn't be damaged as a result of us generating that also meant we just need powerful controls to

sort of tailor or alter the the generated results so not so much just letting her manipulate after the fact as much as it was like sort of guide some of the generations stuff coming into it so here's where we're now so note that these have received receive some art polish that's just part of our standard process which I think can hopefully explain a little bit more of but it's just game captures of some of the islands [Music] sorry cartoons we sped up to make the most of your time [Music] [Music]

yeah so we've got you know initial couple biomes here we've got more in the works these all have a bunch of sub area kind of types sub biomes whatever within that but you know we currently generate all the assets required to make a fully playable Dauntless level so there's no additional work required on the result but to kind of hop in and continue like do full play experience on it and this is where we'll describe a little bit of Houdini so this is our introductory slide so what is Houdini Houdini to us is a visual programming tool not based you have four main types of geometry in Houdini which are points vertices indices primitives and detail you can associate different attributes with

those then you can package your not metrics into assets and you can bring that into unreal this is very similar to a blueprint where you have parameters on individual nodes and you basically promote them to your asset and those show up in unreal in details panel then you have two main scripting languages in knie which is Python index luzon hdk which is the Houdini development kit which we make use of extensively and there's an engine and to ask reading the engine is essentially transport mechanism that brings the data from Houdini to unreal and the other way as well so our technology right now we use Houdini 16 I believe we started on 2014 reading 16 was pretty major update for us because it brought hide fields which we make use of foreign landscape construction our Houdini engine manual plug-in is a fork of funny Unreal Engine plugin provided by side effects which we split off about two

years ago and we maintain our own Fork we'll do our best to point out differences between our plugin and and the kind of current thing forgive us if we wrong you know obviously take official engine docks is kind of gospel over top of any bar examples so right now we're able to generate about 25 different types of actors from Kalinin data and the way we do is is we look for presence of particular attributes on points or pre minutes and we also are able to into pretty much any unreal asset or an Electra or the whole map back in Tahiti for processing as well so the standard approach of inputting data into Virginia is you basically just include geometry but what happens to the data that you're trying to introduce no geometry this is basically what it gets complicated you essentially have to resort to inputting that data as geometry by basic encoding attributes on that and decoding that on queenside so our way to solve this

problem is to convert all the data on inland real disco we do is we include we go through all the properties in the asset we convert the only properties to binary J's and same happens to the if we want to input the actor as well we basically include all the properties of the actor we also Traverse all the components and harvest all the data as well and we do this recursively and on the hidden side we have two operators that we created custom operators one because the binary json into text jason so that we can use Python to basically parse that and extract information and the second one is the actual geometry unpacking we're basically it knows how to fit certain types of actors so any kind of mash these P landscapes and so on so our way of aggregating data is to use a system that we call a recipe system a recipe is an aggregation of parameters

as well as matches and foliage types and landscape materials and let's keep layer objects basically anything that's necessary in order for us to construct and an island so this is an example of a mossy biome forest that was created using our most he recipe so we're gonna quickly cover our overall pipeline at kind of a high level just some contacts to some of the later pieces yeah so what you're seeing on the left left sorry is just sort of a high-level graph that sometimes describes that generally describes the overall connectivity the the main spaces and you know any sort of navigation that's intended so you know node 1 is or area 1 is connected to 2 and you know is implied that there's sort of a clean navigation route from A to B the so this

this is sort of a part of the input that would go into a complete level we take that and we Pro we have an intermediate to D stage where we essentially just use Voronoi a Voronoi fracture cut in two dimensions that we used to resolve a lot of the sort of high-level constraints of the map so that's things like there is a path from 1 to 2 or whatever it is specified it's it's 2d but we are taking some three-dimensional things into account you know like we don't want to slope of more than that exceeds a certain say value so the route from from one area to the other may sort of chicane or cut back and forth and that kind of thing so but there's a lot of stuff that kind of goes into this there's also a ton of attributes that we're sort of calculating at that point to figure out you know how far things are from the island edge groups for this is part of a navigation network this can't be obstructed and and some

outcrops and some other stuff which I'll describe from there we work on a sort of initial like a very clean 3d mesh that describes the general 3d volume we call land massing but the overall 3d shape that has been sort of specified we're not covering a lot of the design tools we just didn't have time to kind of go over a lot of them but we can describe those the thing that's interested just come on up after yeah so the mesh like I said is clean and watertight we deal with like a lot of bowties and another sort of more complex 3d situations we just want to make sure the the squeaky clean results at this point are very predictable for a lot of her next stuff it also toggle the primitive groups here so basically we know where the cliffs are really traversable AR yeah yeah a lot of groups and attributes and stuff like that we define the overall 3d shape like the island bottom is is sort of determined

by sort of manipulating another 3d fracture Voronoi fracture but we're just sort of manipulating the cell points the cutters basically of that to sort of produce sort of a rocky shape that sort of fit our style so the the gif in the top right is just sort of a 2d section cutting through the island that hopefully describes sir how that's created so the points along the top are actually the same Voronoi cells from the previous but you know the ones down the side basically are the new guys so we like I said we've got outcrops it's just sort of our term for this just some land that is sort of shoved up or down on the surface of it these their orientation and some other stuff is determined from some sort of fake stress vectors that just give some initial alignment that makes sense and follows kind of where we were headed artistically but to be honest most of these get sort of shoved

around a particular little shoving around land and the outcrops themselves is the majority of sort of the design input that the iteration that a level creator would kind of go through so usually noodling with this kind of thing is just part of just the exploration from there we determine or we create a whole number of additional sort of assets sort of branch from that point and we instantiate everything from reflection probes to the player starts foliage and fitted landscape that sort of matches well what you just described earlier the overall number of these stages we can't actually describe in any sort of level of detail within a half an hour or whatever that would be meaningful so rather than then covering all of them at like a breakneck pace we decided to sort of dig into a select few

interesting ones and also provide you with sort of a building block and like a starting point so that anyone is interested or whatever knows how to like maybe incorporate a piece of this back in their pipelines so the first example we create what's the point in saying it's Apple it's probably the easiest one you have a cadena network on the right which basically generates random points it assign random transformations to those points we also have an attribute to string attribute on a point that is a path to unreal asset that is going to be in saturated and also have our plug-in specific attributes which defines whether to create instant static meshes or static meshes and then on the Left you have basically unrealistic routing asset that looks that these are none in real they can see all the parameters exposed in detail panel and you can choose the static mesh to instance as well as how many points you want you'll most likely want to provide additional transform information for a lot of things depending on what you're doing so we commonly would have say scale and orient these are just the normal sorry

scale as a flow three orient as a quaternion flow for you yeah it would most likely want to supply those as well to give them a complete transformation for an instance the these are just the standard Houdini wants that refer to the engine docks could you knee engine doc sorry for the kind of better description of what each of those does it's actually a great time to point out that random scattering is actually really the solution for a lot of the things that we wanted it has its place you know it's like Perlin noise it's useful but it's not the be-all and end-all solution for most of the things you'll probably end up wanting to do so quite often we wanted left very very specific arrangement probably an artist controlled arrangement of meshes for a specific thing so we're talking about cliffs here but our solution is actually generic for other situations I'll try

and cover those and the this target needed to evolve so we kind of wanted destructive setup that would allow us to continue improving you know what that would end up looking like and have that just sort of natively work so our approach works off of the idea that if we had a few or a number of samples that described ideal so if art were to arrange in this case cliff rocks in a way that matched what they wanted to see Jen wanted to see we could pick the smartest one that made sense for a new target scenario and adapt it so yeah what we came up with Jen gives everything a really fun name so she called this one quilting what we came up with was we would define inside of a blueprint the ideal arrangements that

covering sort of a range of kind of scenarios and that kind of thing and we could determine a mapping of sort of meshes and transforms you know from from a source scenario to a target sinner so we supply a number of parameters inputs that would help Jen prioritize certain things so this would let us say minimize overall scale changes from from source to target non-uniform scale you know like stretch any along any kind of one axis and overall just distortion differences because the we're dealing with a very organic island we're not we didn't format the stuff from land massing to you know snap to a very specific number of scenarios you know it's not it's not a tile set in that regard so we needed to be you know supply ways of sort of configuring what was important and here yeah you can see some example output of our quilter asset

and this is this is untouched from from the system so this is just the sort of raw output of the biome its recipe and the quilter asset along primitives that we have grouped or marked as cliffs whatever from from some of our earlier things another example of point instancing is foliage in our case we were pumping this into unreal instanced foliage actor and system so yet again Jen could paint flowers and keep adjusting these the way she was before but the result would work with standard instancing and stuff like that so the difference in our plugin and and I think where the current state of the Houdini and Jim plugin is wouldn't be all that different that one change we liked Unreal instanced sorry foliage propagation volume did like the results

but we wanted to take a little bit more direct control over where instances sat and specifically where and how it kind of grew so we just want to add a little bit basically doing reels and the best way to do that was to sort of we rolled the room so we have a very very simple growth simulation I'll describe a little bit more that in a second we use a lightweight like low res 2d vector field it's just noise we call it like a fake prevailing winds and the reason we do that is just so that when we scatter plants and scans sorry plants recede and propagate that we just get slightly more natural looking you know cluster shapes and and bunching so yeah you can see here they're faint but the plants themselves or whatever and the vector field so this is just a quick gif of it showing we run

a fixed number of generations each plant respects like a shade radius of its neighboring plants these are all configurable to recipe which we showed earlier we can get into all the details of the recipe just know that there's a handful of controls like its shade like what it covers where and how it grows so that that could include in water near water in shade or like in cliff shade or stuff like that so we just we gave a few sort of inputs Jen actually controlled like the number of generations that would sort of be propagate to for perf reasons we let every plant try and read properly and if it fails after sort of a couple times or a threshold we cut it out or we don't allow it to try again kind of thing so age is sort of you know tracked along with some of the other parameters and that would drive scale which you can kind of see here so the oldest plants are the biggest winner it's located another example which is our landscape example it's very similar to the point example but high fields - er primitives

so it operates on primitives we created four high fields here we tag them with attributes such as name also with the path to the landscape info object and we create four high fields here one is visibility one is the actual height field and two are the landscape layers we normalize them here as well and on the left is the result I meant to say this earlier but the first few steps with a pipeline involving Houdini engine can be a little intimidating this just there's a lot to take in it's it's a piece of cake for me club but it's not for me so everything you need to theoretically build one of these should be visible or should be covered in this slide ask us if we missed anything but if there's any interest we can definitely supply the assets themselves it's probably worth pointing out as well here sure to cover that earlier we're talking about unreal here

and sorry this the slide specifically would be very kind of unreal but a majority what we're talking about is actually pretty applicable to unity tip as well so so this is the result of our snow we buy them recipe where we had just the water plane level and insurance that affects the landscape layers and enduring effects the foolish instances I forgot to mention that - there's yes a fitted water plane that basically goes along with the biome results whatever from tsunamis we do keep it's worth noting we keep landscape layers normalized ourself on real does have its own waiting within the layer info object sorry but we decided to sort of keep that the normalized state of it sort of track during tracked ourselves just give us more control yeah it was able to much easier to sort of predict or explain exactly why a specific weight ended up

being exactly what it was so the layers in their count the number of layers they're all recipe driven as well so we just showed some sort of normal issue materials when we started with some of that but a lot of the look that we were kind of going for really wanted to give sort of a thickness or a depth to to a lot of the surfaces within the world we wanted a natural-looking build up that you know made all the kind of world objects look like they kind of belong there so you can see a few examples here with snow we did this with sand it's potentially moss with some other things as well dump that one that cumulation and so this was just another landscape layer but this had sort of some special properties that we tied it to an unreal render target and that meant we could do so we did vertex offset sorry Bert vertex displacement essentially or whatever for the vertices and a correction to the normals after the fact

anyone that gets into vertex offset has to bill with that yeah and this simulates just natural build up against those world surfaces I use quotes with simulated a lot because it's not a simulation and we're not doing any Houdini's capable of also it's a really cool simulation but this is not a simulation really the goal here was very configurable art configurable profiles and and we really wanted Jen to be able to kind of kind of drive the result and being able to control like its max depth and that kind of thing or whatever or where and how it built up or whatever is just anyone that works a simulation knows it's a little hard to kind of tame those sometimes I make sure that the results for those are predictable not that it that route is impossible it's just that we opted to drive that a little more carefully so here you can see it capture for the engine as well this just shows like in the case of trees the the and sort of leeward profiles that you

know let Jen basically kind of build up a shape that she liked for natural-looking snow buildup yeah I'm glossing over a number of the details with this we're not trying to protect anything or like just not give those details that just didn't all fit but we're not trying to hide anything so if there's anything that's sort of useful or ever feel free to ask us or definitely want to fill in the details so things so we kind of basica to describing things that we've learned I've worked in this the first one is just don't do the goober acid it's hard to maintain in debug then what really worked well for us was decomposing all our pipeline into multiple stages so we have about 50 necessarily for sale and they got rearranged a few times so the modular nature of how did pay off so that allowed us to work in parallel on some of them as well as allows us to switch to certain parts of the pipeline out or disable in order to speed up the cook and also allows us to achieve some poram polymorphism as long as you respect the same person tuples also

represent that and use Python for that so that's basically what we do and then we optimize that into either of X nodes or hdk and speaking of hdk it was extremely useful for us to actually be like spin up the visual studio which is basically attached to energy K node and debug that because that was just easier way to debug geometry down and do printf debugging yeah you can put a Python snippet into you know any of these networks super simply and super powerful there's a lot you can do with it it's not quite as fast as as Beck's or HD K but we found situations where it can become unwieldly especially when you start if you're shoving like you know three pages worth of stuff into a wrangle you know you're getting to a part where I found it just kind of hard to deal with this the solutions to some of that but for just the best debug ability vehicle and I are both I think a little bit more comfortable in a kind of a sea world so that made us kind of opt for HD K in those cases as well so another thing was ability to save the

hip file so basically something went over to our system our artists would save the hip file and we would open that scene in academia and just inspect that and debug that the like we mentioned before getting a very unreal compatible output ended up really being I feel like a good choice we did opt for a layer to discuss options like not using landscape and some things like that briefly but being able to just resume editing one of these levels and exactly how it was created by hand or whatever was super valuable and you might rephrase some of this is like getting us to kind of make sure we were thinking more unreal like in some situations versus getting Jen to say to think more Houdini like most of the Houdini stuff is essentially invisible or not invisible not a great word but we did not ask Jen to sort of get more Houdini proficient basically in this so and the last point is being able to input arbitrary types not just the geometry back into academia was

extremely valuable to us it actually allowed us to build assistant do determine is decreasing essentially yeah so like raw proper classes back and forth is it's really nice whereas the attributes themselves if you try and contain those any other way it gets real messy so yeah thank you [Applause]