Embracing Animation Complexity in EA UFC 2 Through Procedural Techniques

[Music] [Applause] well thanks everyone for coming I originally wanted a much longer title for our talk but mike-mike nixed it so I can see why okay introductions quick introductions so I'm Jeff Hara where I'm senior software engineer at a Vancouver currently lead gameplay engineer on UFC I've been working in the industry since 2000 so I've been working at EA for 16 years which along with my increasingly cleaning hair makes me feel very very old and I'm Jerry Matheson lead gameplay animator at EA Sports in the UFC two title film industry since like 1999 store that made for attainment making kids grow tunes and I'm here making me do the games alright so here's an outline of what we're gonna be talking

about today so I'm gonna go first and I'll be talking about all the technical engineering stuff I'll give you an overview of our dynamic grappling system kind of the the design change that we went through from UFC 1 to UFC 2 and talk about how that influenced our animation requirements I'm going to do a bit of a technical deep dive on the relative a case system I talked about that a little bit last year looking back in my talk I don't think I explained it very well and there's a lot more depth that could have gone into so I'm going to dive into that because it's kind of a fundamental building block to the procedural system we put together this year and then I'll go over the what we called procedural mocap that's a term that Jeremy coined after seeing how this stuff worked how we solved our grappling problems for this year and I'm going to talk about essentially how one gameplan tomater can create that much content for a feature I'm going to start with UFC one you can see doing it the old way how much content we created and then on UFC to how we improved on

that and a big key part of that is just procedural techniques that kind of a us in a big way on that and that I'll come to a conclusion where we've arrived so it's a big exciting week for us our game launches tomorrow so you're not familiar with our game or the UFC here's a hype trailer to get you all excited for the talk the notorious Conor McGregor black trunks for the champion Jose Aldo jr. Conor relaxed and smiling [Applause] [Applause]

all right there we go what's fun Thanks okay so one of the big back of the box features for this year's game is something that we called dynamic grappling so to understand what that

system is I'll take you back to UFC one and talk about the grappling system there so it's pretty pretty straightforward kind of stuff so when you're when you take the guy to the ground you start off in some kind of Idol depending on what position you're in the top fighter has a bunch of things he can do the bottom fighter has a bunch of things he can do let's simplify it a little bit and say the top guy can go left to right and the bottom guy can go left to right so if the top I presses to the left he'll play an animation of him doing something to the left that could be a submission the transition attempt in this case he's trying to pass the bottom guy's guard by pressing down on the knee so you've got a start animation to initiate the action you've got an idle pose where the game kind of figures at the transition logic and then you have a cancel animation to bring you back from where you came from and a finish animation that executes the technique and takes you to where you're going similarly the guy on the bottom presses

to the right he'll play some move in that direction similar similar thing a submission transition something like that so what happens if the top guy presses to the left and the bottom guy presses to the left at the exact same time so in UFC one the top guy pressed to the left he would play his transition animation but if the bottom guy also pressed to the right and the game registered the input just a little one frame later he would get absolutely nothing no response from the game that he had done anything he'd be locked out of any transitions the only thing he could do is recognize that the top fighter had started to transition and play a denial animation which just prevents him from doing his action so this wasn't very good for users they felt frustrated oftentimes they'd be pressing on the stick and get no response they thought the game was broken its unresponsive I don't understand how to play

just generally a bad bad experience for people so in an ideal world we want something like this where top guy presses to the left and he gets an action bottom guy presses right he also gets an action so the two players could act independent of each other playing an animation that correspond to what they were trying to do no matter what the other guy was trying to do increase responsiveness and key increased that sense of engagement with the game so let's get a sense of what kind of numbers we're looking at in order to support a feature like that so just in full guard this position here let's say the top guy has eight possible transitions he could do eight transitions submissions whatever so left-right up-down and then holding a modifier left-right up-down and the bottom guy also has a transition say for example a similar deal this would represent all the different idle states

that the players could get into in full guard it's not 100% accurate it's this kind of the back of the napkin kind of calculations we were doing at the beginning to see what we were up against to try and implement this so you're looking at eight times eight 64 different states I'll be like top guys going on the right bottom guys go on the left top guys don't know the right bottom guys going down all those permutations right so in order to figure out what kind of animations we need we need four different animations for each one of those states we need something that takes you to that state an idle animation once you're in that state I cancel that brings you back to the state you came from and then a success animation that takes you to wherever you're going from each of those states so we multiply by four and then the game has 15 different ground positions so assuming every ground position has the same level of complexity asgard which they don't but some are closed some are bigger some are smaller we've got to

multiply that number by 15 so we got a really big number here to put that into context this little piece that I cut out at the bottom that's the number of animations that we built for the entire ground game in UFC one and the big picture is what we would need to build in order to support this system for UFC two with the same budget the same timeline same number of animators that kind of stuff so on the face of it it looks absolutely impossible unless we can come up with some kind of procedural system or something just some way to optimize this so that we can get like a tenfold increase in content creation not trivial at all so before I dive into how we did that I want to talk about relative that case so relative ikk existed in EA MMA like four years ago existed in EA UFC it's used by a bunch of different games in da sports it's a system primarily designed for playing two player

animations on characters of different scales that was kind of its first purpose and it's taken on a different meaning this year for us I talked about a little bit last year but I'm going to dive into some details same rules apply I'm not allowed to show any of the editors in our tool but I kind of mocked them up a little bit in PowerPoint to give you an idea of how it works so yeah two player animations let's take a look at a concrete example here so in this one the bottom fighter is reaching up and grabbing the top fighters head and then canceling back into the idle so what does that look like if we make the bottom fighters arm comically large obviously the contacts point contact points don't line up the animation completely breaks so with the relative I K system what we do is we offer our constraint between the left hand of the bottom fighter and the head of the top fighter and we bake out the position of

the hand relative to the head into a new channel in the rig for every frame of that animation this is not how it works on the other hood but conceptually think of it as a bone that's parented to the head of the top fighter in the top fighters rig that's just constrained to the hand say a motion builder and then baked out with that global position right now you might be thinking how how are you gonna generalize that there's like the hand could be relative the head to the arm any number of joints on the other characters skeleton and vice versa are you going to put that many joints in the skeleton one of our really clever clever engineers Dominick Parker figured out a way of encoding that information in the rig so that we could do it with a limited number of slots and it would blend across animations that's kind of the secret sauce of this solution but if you have like a limited number of joints you want to make relative to each other the bone approach is pretty much exactly what we're doing here so we baked out the path of the hand relative the head all throughout the animation and the animators can offer a

curve that defines how tightly the hand is following that constraint so it can become relative the head it can kind of ramp in as it starts approaching the head we add a hundred percent when he's holding the head and then ramp back down to zero as he lets go so when we play back this long-armed character on the same animation we now have a target point for the hand relative to the guy's head we just snap it to there using AI K and the way we go so this is what that simple example looks like once you turn on the relative that K system pretty decent results so that works well for when the arm gets really long and you need to figure out where to put it what happens if the top fighters head gets really really big that same approach doesn't work so what we did this is new for UFC two in the talk last year I was saying that we wanted to move our I K system into the

physics system because we're having a lot of fighting between the AI cave and physics one would be pulling in one direction I could be pulling in the other and you never really resolve anything we accomplished that and it gave us some new tools to use when trying to calculate these constraints so what we do now is we cast array from the head to the hand and figure out where that ray intersects with the collision volumes and then we use those collision points on the surface of the mesh to define the relative constraints so now as the head gets bigger the relative contact point moves out depending on the arbitrary size and shape of the collision geometry that's underneath so let's look at that in action if I grow the top guy's head and grow the bottom guys arm you'll see the results that you'd expect the animation played back

so it's a simple simple example but in our game we've got animations like this where the characters are all tangled up like spaghetti so someone has to go in and offer all those constraints now that was a huge time consuming process in UFC one one of our engineers John too wrote offline batch process to go through and analyze all the motion figure out which joints were in close proximity to others and where they settled and came up with a first pass of all the constraint offering on these and then sometimes they would work sometimes they wouldn't animators would have to go through and tweak them in some cases just a little bit in other cases a lot but we were able to speed up the process by automating that stuff and every single animation in the game has a set of these constraints built into it that's kind of the the base set of building blocks that I had to work with when putting together this procedural solution now I've talked about using this for scale compensation

but it can also be used for other techniques one that we use often in UFC one is animating against static punching dummy so you gotta have a guy on top of another guy and the bottom guys animation is perfectly still there's nothing animating there and the other guy would just punch at his head grab his arm or whatever and then you could play that animation against any animation for the bottom guy and it would just track and move appropriately so it was a good way of like kind of reusing animations and two-player context and that type of technique is where we're going to go to with this time a dynamic grappling system it also worked on the feet now I need a short volunteer Tim do you mind coming up here we were doing a rehearsal last night and we know I know something in the video that's coming up that I didn't really explain so I don't have any slides for this and Tim was begging for us to do something some sort of physical comedy because we did it last year so I'm gonna use Tim as my as my get closer come on you gotta get a camera all right so this

in the next video that I show you keep an eye out for this pose here so this is the Muay Thai clinch now obviously I can do it - Tim no problem right now try and do it to me he's got to reach way up not easy to do right so we added a system where we could set constraints between arbitrary joints same as in relative okay but only on one axis so you could define what access it it applies to so if we're in the same pose here I can apply a constraint between our hips and I came my hips to his hips but only in the y-axis so I can drop down like this and now similarly if Tim's trying to do it to me I can drop my hips down and now he can do every every move that he wants every I can do every move that I want against him now obviously if Tim's doing it to me he can't ika his hips up right it's not possible so what we did was we created conditional constraints where the condition could be the height difference between the fighters so those

are a couple cool little tricks that we used in this system Thanks to everyone round of applause for thing so here's an example of this system working in in-game in UFC - with one of the most extreme height differences that we have this is Stefan Struve who's I think 7 feet tall or 6/7 um yeah close to 7 against Mike Tyson not sure how tall he is buddy's pretty short so you see the animations hold up pretty well here he takes us back pretty soon he'll get to that same tyke Lynch position that I was demoing with Tim or a almost next one right here so now look at this pose look at this animation these are animated against same scale characters and then Mike Tyson's going to reverse it soon and get the Thai clinch on Stefan Struve that's the exact same animation playing back and I can't believe that this works it holds up

really really well I don't even think this combination was tested by QA we saw it like last week we're like oh crap and then it actually worked so our clinch engineer Dominic spent a lot of time handed offering that stuff and then just for fun here's another way back in a way I'm not gonna let you near me at first all the ranges Hey that's an example of another extreme hide districts that we weren't expecting to have to support in the game and really really tall hair that can really mess with contact points anyway let's back to the problem that we're trying to solve a huge complex animation tree that we need to build and

not enough time or manpower to do it so let's dig into some concrete examples here this is an example that inspired the the technique that we used to solve this problem I was looking at this animation so the bottom guy just puts his hand on his knee that's it very simple and then what do I want to mix that with with the top guy top guys putting his hand on the bottom guy's knee it feels to me that those should be able to play independent of each other without a whole lot of animation work because the important things are moving independent of each other you look at them side by side there's not a whole lot of overlap in in the important movement in both the start and the cancel so in order to build this prototype I asked Jeremy to throw together something that we call a mixed idle so a mixed idle is a just a regular hand authored cycle that is what they should look like once they get to this state of one guy going to the right and the other guy going to the left so this

is what Jeremy put together for me there's a lot of these but it's not that exponential number in that first slide this is a manageable number of animations this is what I need from him in order to even start on some sort of solution so this is what we need we need we go from the base idle to top guy going to the left base out or bottom guy going to the right and then if the bottom guy is already going the right top guy go on the left and vice versa so the green and green arrows are the animations that we have from UFC one to work with and the blue arrows are the animations that we need to figure out a way to produce in a very very cheap kind of way so we need to go from the idle on the left to the Auto on the right let's focus on that example here's the animation of the bomb guy putting his hand on the knee that seems very important so let's make that base so what I'm doing here is I'm building a set of layers I know that I need this

animation because that's the piece that's missing so I put that as the base layer now he's not coming from the right pose or going to the right pose in this base layer so I'm going to toss in the from idle and the to idle into my set of layers here and now I'm just going to start with the from idle blend it out and end with my two idols so I'll blend that in so I got a bunch of curves that I can now manipulate controlling the blend to him from the idle and with that setup you get this kind of result so we start at the right pose what we're doing we play the animation on the right hand that we want but the top guy's hand pops to the chest and then back to the leg so if you focus just on the right hand at the bottom fighter everything looks pretty good the only problem is the top fighter are kind of popping around as he blends out to the transition animation and back in again so I toss in two more idles the from idle again the two idle

again I give the ability to mask in and out different parts of the body on those layers give another set of curves to play around with surprisingly this this next iteration was enough for us to solve all the problems there was enough flexibility in this setup to get everything that we needed but I didn't know it at the time okay so now we change our mask I want the hand animation from the base layer so I mask that out that's what that little red dot means I mask in everything else for the bottom fighter and the top fighter or sorry just for the bottom fighter and so I want everything from the bottom fighter except for the right arm so right arms coming from the base layer everything else is coming from these from into items I just do a straight crossfade from the front and two idols for everything except the right arm and then similarly for the top fighter I just do a crossfade from the source idle to the destination idle because he's not

doing anything interesting with his body so we just there's subtle changes in the idle but we just do a crossfade there and this is the result that we get when we do that kind of layering hey that looks good I can work with that it's got me really excited I think we've solved the problem let's try some more examples just to see what happens and there's at this point everything's hand offered all these curves are hand offered here's an example where there's quite a bit more motion in the body of the bottom fighter but the curve setup and the masking setup is surprisingly similar to the first one so using that technique we get a result like this it doesn't look like super great but it's passable we could put this in the game I could ship with this if I have to but it could certainly look a lot better so now I'm starting to get a little nervous this might work but it might not work let's see where it goes from here just as a as a test I got Jeremy to take this into motion builder and clean it up and see what he came up with and timing but not spend a lot of

time on it because I was like okay if we have to tweak some of these by hand how much work are we looking at so the one on the right is the tweaked animation that Jeremy did the one on the left is the fully procedurally generated one the one on the right looks better but it didn't take a lot of time for him to put that one together and the improvement minimal the timing looks better it's not quite as loopy but yeah we're still in a good spot here here's another example that came later in production when we realized that we could use this for strikes too so this is something that the bottom fighter can do he can punch the top fighter in the head while they're addling here and I wanted him to be able to punch the guy in the head while the top guy was in the middle of doing something so that you get that continuity of punching no matter what the top guys doing so we put this animation in the same system instead of the from idle to idle setup it was all just the idle that I want to use all the way through take the arm and a little

bit of a torso in the head from the punch the rest is just coming from the idol that they want to be in and this is the results that we get from that that looks awesome that totally works I can work with that as well so this is starting to look up there were some examples that didn't go so well any time the guys were kind of on their feet there was no way for me to get any kind of footwork and no matter what I did with the curves tweaked them didn't matter it basically ended up like a really crappy looking look where the head passes through the hips and it's very slight E and linear and this we kind of ship with this absolutely impossible to ship the game of this so what we ended up with in the end was a spectrum of animations so the the the masking and layers and curve setup that I showed you that that was final that's

what we ended up using we were able to use a the same pattern for so many transitions with that looping all bones except for these and asking out this but they were all kind of different masks and different layers so what we did was we put together a process where we could identify the relative I cake constraints that were the same between the source and animation era source and destination idols and procedurally generate these curves and masks as a first pass and that worked for a large large number of animations but then there were other ones where we had to tweak some of the curves try and get a little bit better results and then there were other ones where tweaking the curves wasn't enough so we built a pipeline that would bake out the resulting animation into a an animation that could be edit in MotionBuilder it's kind of like a one-button export straight into motion builder the animator could spend an hour cleaning it up it would go straight back into game already hooked up and finally we have some that were just so bad that we had to use traditional mocap so we had a spectrum here it I was aiming for fully

procedural because you get all the memory savings from that obviously a lot less work but we were sitting in a place where we knew not everything was going to work that way what we didn't know is how many would fall in which part of the spectrum so we had the following concerns coming at a creep a huge memory concerns because if we have to do a tenfold increase in animations without increasing our memory budget we just can't do this feature the other one was the quality concerns how many are going to be on the mocap side how many are going to be on this fully procedural side and are we going to be able to create enough content to support this are one of our brilliant engineers Sandberg Avanti put together an animation streaming solution for us so I'm not going to go in a ton of detail here but because of the rigid nature of our state machine if you're idling in state a you have all those red lines already loaded in memory if the guy presses on the stick you can just play that animation because it's resident in memory as soon as you play an animation that transitions you from A to B we

unload all the animations that come out of state a start loading all the animations that come out of state B and as long as you can load all those animations in the same amount of time that it takes you to playback the animation from A to B you're golden and he got the system working so now we're sitting in a place where memory is not of concern and we have a system that will scale up ridiculously big for future iterations of the game so super excited about that memory is no longer a concern the big concern was the quality we had no idea if we could get this done on time or to a quality level that was necessary to ship the game so we kind of took a leap of faith here there was nothing else we could do we mitigated risk by focusing on workflow efficiencies and prioritizing the work just in case we ran at a time and coming up with some backup plans for some of the more fringe positions in the game so that led us into production big unknowns not really sure if this is going to work but a plan in place that kind of feels

like maybe it would and Jeremy will take you through production hello okay so next we're going to talk about the animation side of the a Sports UFC - and I think the best place start is UFC one so I want to turn it back to UFC one and talk about our approach to workflow there well I talked about the early design and workflow approach I want to talk about the changes we faced and the challenges we faced sorry and post-mortem so UFC one the way we've you know faced the feature was the way that anybody does I meet with Liam Miller our lead game designer and we break down that feature and see kind of what he had in mind next I would collect any visual reference that would helps the defies

feature in my mind and make sure that's the same page as everybody else there we go and then what we do often is either use a piece of paper you say camera we go in and we would break down the individual parts of that feature this is example of a transition on the ground all our transitions for the most part have five parts here you see Jeff and Liam are starting in the full guard position if the player were to press write the stick they would give start animation then they would enter a ready animation and depending on their inputs they would have to get a success a deny or a castle animation so now knowing that each transition has five parts and how many transitions I can then create a shot list for motion capture studio on

the bottom right you'll see a pose library there that particular pose library has 87 pages with several poses on each page that's not even all our poses that that's our Bible when we go in that's a visual reference to the actors and then we go in to the studio we capture the data if you look on the right hand screen Liam Miller is weighing down the cage and beside him is that pose library that thing's essential then using proprietary software I'll go back to my desk and I will painstakingly and tediously for that data it's not a fun process and then for UFC one we would clean it up in-house this is Dimitri Cooper in one of our senior animators cleaning up a grapple transition and then we'd review and read

revise revise to a shipful quality this animation is still in the revision stage but it's it's in good shape it's it's getting there so why do we do it this way we had the time when we had the head count we had a head count of one and we figured that would be enough for this feature and I was worried that the complexity of mixed martial arts and jiu-jitsu would be lost in translation using strolled development the sport is very technical our fans are very knowledgeable and they would know if we messed up grips have to be just so posing just has to be just so I was wrong in this assumption for UFC two we did use a shield development and it was to a huge success if you guys want to hear more specifically how we did that say GU Tanaka and I did a presentation at XG s last July and Vancouver so if you want to find a little more about that

so working that way we end up a six hundred two player grappling nations for UFC one the problems we face working this way our name Convention discipline was lacking some animators named for example our half guard position half guard others named it half G some people named it start the start animation of the success or others attempt neither are wrong but they are consistent and that made a reputation less effective when searching out these particular assets problems like that and other problems completely different led to overtime led to late implementation which led to poor testing poor tuning I skip something there no so post mortem coming out of that we decided right away we

definitely did tighter named prevention because of the animation number was limited this made the ground game feel linear limited it was turn-based and visual feedback was lacking we did up with a 7 out of 10 on Metacritic we really wanted that 8 out of 10 we needed a bigger game we needed more animations and I need to embrace external development so moving into UFC two's production cycle I want to talk about the prototype of the game preparing for production and then production itself so the prototype a early entry probe Jeff came to me and he said hey Jer I got something to show ya and I came over to his desk and he showed me his version of the ground game very early on and it looked smooth and

looked organic looked real and I was really exciting I said how did you get this he said well I have this nifty new grapple controller tool and he said I just want to show you the decay ideas coming out of this tool so he had me mock-up a mixed Idol which I'll explain more about later and he put through to existing UFC one assets and connects them together and what we got was this kind of weird programmer art linear animation and I looked at it I'm like that's interesting I looked at it again and I thought well our mocap isn't that great either and this data comes with the begin pose in the end pose already imprinted and perfect and all of he was clamped in between poses and I like that's actually kind of really cool so I was like let's let's see where this goes i like this so far so we can try to break it down like what if we did the old way and we would have to do a law of our fsz all the way but if they do everything the old way you know I still be the Liam or Jeff and

figure out how it's all going to work I love to gather my visual reference for the motion capture make sure we're on the same page I still have to create these mix titles which I'll talk about more later break down the design figure out that name convention super-important create the shot list not fun it takes a long time go and shoot the data can be fun not always fun or the tool you know or the data which always sucks no one likes doing it and if you did deal way we clean up in-house which we wouldn't this time we'll use actual development some of it in-house and external and then review and revised with Schiphol quality so we knew we couldn't do the whole game that way so what if we use Jeff's tool so for the procedural data we've maker mix titles we throw it through Jeff's nifty tool

we'd press export ideally we ship up ship it off to a technical in India and you know we're good so we decided we're going to try that let's prepare for production let's see how we're gonna do this all so I'm going to first talk about the mix titles we all know what that means on the Left you'll see our regular full guard position if the top man presses left on the stick he gets this idle now if we go back to our default guard idle if the bottom and press is left on his stick we get sorry the bottom model there in the middle and then if they both press left in the stick you'll see here in the full guard position the top man's right elbow you'll see here at his Posey's dig into his elbow into the bottom man's left leg you'll see again in the far left picture the bottom man's left wrist is now going

for the arm drag it's around the top man's left tricep and in our mixed idle when they both pass left to stick the top man's right elbow is dig into a bottom man's leg well at the same time the bottom man's left wrist and is gripping the whoops gripping the top masked tricep perfect that's what we need so we now had to figure out what kind of data was had come out of this tool so we knew how to best prepare for that we did like a vertical slice of the guard position we did some quick math out of that so we figured based on the guard position that 80% of our data would be easy to process it would take about 1.5 man hours to clean up each little transition 10% the data would be difficult to process going back to the easy to process that would also be great as said to Arsenal development it's easy for them difficult to process is 10% that would be 3 man-hours and those would be awkward to explain to show

developments we thought let's just keep that in house the small number the perfect keep in house and then ten percent would be cannot process we'll have to go through that whole motion capture process and then once that's done be able four hours for a person Institute clean it up so just an example of easy to process data this was the kind of data we're getting for the most part for the tool the bottom man's left arm is going to top man's or straight above man's right arm is going through the top man's left arm and the bottom man's man's legs are crossing through each other it's very linear it's it's not bad it's definitely something we can work with but it's not perfect as you can see so that's about 1.5 hours for first cleanup difficult to process data this is a bit trickier it looks kind of simple at first but the top man's right arm is going through the bottom man's left arm so we're gonna have to explain to a person you got to

bring that wrist all the way to his hip to pull it through we'll probably have to separate the two the top mount to push off a bit separating their bodies and then with the pin the legs rather try to explain how to keep it in-house keep it simple and then difficult to process data as you can see the top man is like teleporting across the mat he's sliding he has no weight he's like he's on the ice rink we probably could spend some time clean that up but why bother let's keep our visual quality high let's just go reshoot this so movie into production this is what a production was going to look like for all the types of data Kalfas tool we were gonna walk to our name convention create our mixed titles we press the export button we would you know take all that data and we would estimate how long it would take to do it all with makers Docs and spreadsheets so we know where it was going to be at what time and what state

we ideally ship off the ease to process data to India nice packages from the cleanup difficult to process data would keep in-house clean up and then can I post this data we would just go in sure in our capture studio and this is here's an example of some of the naming convention we had to work with it's the longest name convention have ever worked with this one it's clearly UFC to the G means it's the ground trans means is a transitional animation the sub means it's the submissive the bottom man who is initiating the transition back mount half guard means it's going from back mount to half guard right means based off the top man if his head is north and his groin is south is going counterclockwise and Jeff can explain to the last four parts of the name yes success means he's actually succeeding right nations transitioning submission down advanced right none none those are

kind of the states in the state machine so the state in the state machine is defined by three things what grappling position you're in what slot the top fighter is in and what slot the bottom fighters in though that Triplett defines a state in the state machine so those last four pieces are kind of giving all the information to know we're coming from submission down advanced right and going to transit none trans none from the back mount state to the half guard state and the important part here is that this does mean something of somebody and sees this or children find these assets which is huge is this number of assets so this was a big win for us so now it was important to keep that difficult to process data down to a small number and I explained why we had 40 default poses that's just your full guard half guard anybody watches UFC you know these positions we had 120 to neutral poses these like these mixed idols he's ready idols we

had two or sixty one submission poses from our grapple transition state going into our submission state that's a total of four hundred twenty three potential poses that you're gonna have to you know you're writing out your your shot list your figuring out what goes where you're doing the name convention you have to make sure you have all these poses prints out to show your actor this is just a small example of some of the poses that the actor in the data match the closer it matches these for your external development team to use you then have to go and take that data and put into motion blur file and like provide the sources with the exact start pose a tenpo is just a huge nightmare so to keep that number small was crucial so the end result of working this way UFC one had six hundred two player ground animations UFC two had over 1957

two playground animations UFC two's and tug grappling system had 3153 two player animations touristy three of those were procedural we ran through Jeff's tool and were untouched by an animator 4.37 were procedural animations exported and processed being an animator touched them 98 we could not it was data that was was unusable from Jeff tool so he had to go and actually re capture those and clean them up the old-school way and then one thousand three of seventeen transitions we actually went captured and cleaned up in our capture lab that's UFC want to you see to combine at one point I want to make about this to is even though in the end we ended up motion capturing a bunch

and cleaning up a bunch while we were in early production in pre-pro we had a template for almost all the transitions in the game so we could actually prove out the design we could convince people that this was worthwhile from a gameplay standpoint even if the animations weren't perfect or beautiful they were a hundred percent functional and it gave us a template to work with all through production if we hadn't had that early prototype no way anyone would have signed off on this feature to begin with so the procedural aspect had huge benefits in pre-pro even if those benefits started to dwindle as we got through production and everything yeah it's really worked really early on use the procedural data which is huge um so 587 transitions we said detect occur in UFC 2 which was a huge win for us 80 of those were of just tulle this number could have been much bigger I spent so much time managing my ex dev group that I would have these small gaps in time where I could work on actual

data cleanup processing and so I like those animations because it would just be like an hour hour and a half and I could I would have to stop halfway through and go back to managing so I like that but this guy data is perfect for your show the vote partners if you were to use that system and so yeah if we did this the old way without Geoff's help this would have taken essentially three animators in-house to do and that just was not an option so just visual reference if you compare the amount of data we created UFC 1 T UFC 2 UFC 1 being the blue UFC to be in the red you can see it's like over three times the amount with the aid of Jeff tool and just here to break down of the type of data we ended up having so one left it's our motion capture and then the two small pieces is the data from Jeff's tool and you can see the amount the top

wedge is procedurally edited animator touched at the bottom wedge is untouched by Jeff tool which is amazing and then of the data Jeff's will produced you can see on the left that's the amount touched by an animator and then on the far right untouched and then that top wedge is the amount we couldn't use we had to go in to capture so in conclusion I think UFC 2 was a success success because of our organizational protocols all that docking knowing what asset was weird in the world and what state we had to know that over time and what's data was in the game often we just put in raw mocap data for Jeff to use we get all set up and then we have no when it was old and I'm here for India we've know we were placed with new data and the intervention if I didn't mention already and then it works drill development was a big part of why we got this done and then one of the biggest pieces was Jeff's tool for sure it was a

massive win and we ended up with our ADA attend last time I checked which is very happy he might have gone off might have gone up it's 81 that was to say that and so in conclusion a lot of you I'm assuming watch house of cards there is a scene in the house of cards where there's all these monks in the White House and they're sitting over this table and there it looks like they're kind of grading chalk on to a table creating this pattern it looks like painstaking work it looks not innocent tedious and it looks long and hard when Jeff came to me and showed me the early prototype and I kind of understood it would take I knew it wouldn't be sexy work like overwatch but I believed so much in the result and as you see the result of this you know amongst works this beautiful inspiring pattern that is

pretty mind-blowing and I just saw the bigger pixel tunnel like this working with always little pieces and doing this track and documents and working with X dev it's not sexy work but I had the big picture of mine and I was super excited I love this for mixed martial arts I love grappling and I was like what if we could actually make the first legit grappling feature in a game and so that's what you know that's why I went through all the hard work was for that end result I was so excited to see it and here's an example of all that hard work [Applause] powerful right hand break possession mountain here for submission

he's got that arm bar lit off right here [Applause] beautiful big diving punch to the head Maya's right eye is looking pretty swore he's doing a great job of moving and transitioning here on the ground could elbow to the head five major babies inside control again so that's literally thousands of animations from here I'm gonna hand it back to Jeff I want to go through a few of the learnings that I took away from this project so we're in pre-pro we've got this seemingly insurmountable problem unless we can come up with a procedural solution and procedural is sexy I could probably do a GDC talk on this awesome procedural solution I really want this feature to work we've got to find a way to do it so I start on the procedural solution I'm

starting to get these mixed results and it's like I don't know if this is gonna work there was a decision point there where I could double down on making it a hundred percent procedural and trying to introduce our physics engine into this offline animation generation process of some of those clipping issues would go away but that might introduce more problems I could maybe try to add more intelligence in the auto matching algorithm to try and get better results but don't lose sight of the forest for the trees the goal was actually to ship this feature and figure out how to make this much content so put the cool technology to the side and focus on workflows focus on tightening up the export process make everything as quick and as fast as possible and figure out a way to work with the animators to get to the end result that was a big one another one so game industry is kind of material a little bit over the years animation has matured of kind of falling behind or

trailing behind a little bit but we've got a bunch of mature tools most engines have like some pretty good generic state machine technology animation blending technology all that kind of stuff so we've got some really good general purpose tools but when we scaled up to this or we scaled up to this volume of content those generalized tools just weren't going to cut it so there's a tipping point somewhere where the number of assets that you need to create inside of a system it's not worth using those general purpose tools you have to build something very specific to the problem at hand so that you can extract every tiny little bit of efficiency out of that process because it's going to just explode in in volume and every little bit counts so we spent a lot of time we wrote a custom state machine for this I'm annually entered all 4,000 nodes in that state machine but it contained all the metadata that was needed in order to

understand where it fit in its place we could write a bunch of automation tools off of that state machine all of the procedural transitions and animations were built off of that base state machine layer so everything kind of came from one central hub and the workflows for offering that state machine were 100% designed just for what I had to do with that state machine and if we hadn't done it that way there was absolutely no way we live have had time to do this it was a massive massive undertaking and that initial authoring of the state machine could not be automated in any way because it had to follow the real life sport of mixed martial arts it wasn't a homogeneous system that translated across all grappled positions every position had a unique flavor and difference based on the what arm was grabbing what what could you do well another guy was doing a thing and it took a lot of knowledge of the sport and custom tools in order to put something together that complex and this was a big one from this is what I wanted in my original title but we cut it out to quote Nate Diaz don't be scared homie

we were facing something that seemed insurmountable this is probably gonna be really bad advice to give to you I'm gonna do it anyway because it might be inspirational or it might tank your next project I don't know we were kind of looking at this massive amount of stuff that you have to do and huge unknowns and you've got management asking you to come up with estimates and then you've got to break that down into who's going to do what and honestly like there's a point where I don't know the answer like I cannot estimate how long this was going to take I have no clue so I choose I get to choose are we taking this risk or not because people are trusting my judgment to decide whether we go forward if I was conservative I would have said no we can't do this because there's too many unknowns but Jeremy and I were like super excited about doing this it was like [ __ ] it let's do it right or die and so yeah just don't be don't be scary if you're really passionate and you make solid contingency plans and you know you have decision points on the product in

the production cycle where you can take a left turn or a right turn if things don't work out the way you want just go for it you're working with a team of really smart people and you'll be able to pivot when you need to so yeah just don't be scared to take risks awesome things can come with it probably bad things can come - if I'm not gonna talk about and that's it we've got a few minutes for questions if anyone has [Applause] over here hello I'm John Butterfield I have a question if you ever went back and edited any of your mix titles excuse me like so that if your start and end poses changed how did you realign those throughout the process so there's two pieces to that probably what Jeremy's going to say but at the beginning there was a lot of editing of those mix titles so we would try and keep things procedural for as long as possible to kind of work out all the details with those mixed Auto poses if

we were to bake something out and have it tweaked all of that procedural information was still stored in the data library so we could always flip back to the procedural one regenerate it and then do the cleanup again so that's my answer that's pretty much my answer there were times where if it was a small change I just go into the edited data and just left the pose on they polish it further but sometimes you just regenerated know a bunch of so it was kind of think you know grad work at that point the bigger problem was with if any of the transitions that were used as a base for the procedural generation changed timings that was a much bigger challenge because when we oughta generated these transitions they would reuse all the tagging information from the animation they would reuse all the relative ik' information from the original so no one had to manage or maintain any of that tagging or or curves or anything like that as soon as you baked it out it basically becomes its own entity where any changes to the timing all those tags have to be updated we also had some problems with some of

the auto curves have to be regenerated if the base animation changed lengths and stuff like that so that's a tricky thing to manage for sure okay thank you Jeremy when you first saw the procedural system you said something along the lines of well I'm okay isn't that great anyway could you elaborate on some of the problems or mocap sees where this procedural solution looked like a better way to go I would know like it wasn't like it was better because the data was good or bad it was when you go into the motion capture studio and you get these actors posing themselves on the ground and real weight and the difference in their bodies this had a challenge sometimes mashing those poses or we had the problem with UFC one we were kind of went to white belts making jiu-jitsu and we kind of had some bad poses and we had these black belts coming in with better poses and their hips be more on the side they'd be like using more of a deep half kind of position and then we send that

data off to India and they're like hey these poses don't match your old poses and making our work quite a lot harder for us so with having Jeff's tool just that was the biggest part of it was just baking out those poses and then when I take that data and I packed it up said to India I don't have telling you anything just clean up the middle whereas when I sent them the Moke motion capture data I'd have to go and find those poses and when each in each package each transition have to give them each start and end pose just to make sure the only things as simple as possible and there's a lot of front-end work on that and you get rid that friend work is huge I don't know that people will really appreciate how complex a state machine with those four thousand transitions is when you go into mocap and like you got your 8a mocap shoot and you're aggressively trying to capture like 250 moves because we need these moves otherwise there's all these massive holes in the state machine you're going to be like on top of every single animation that you're trying to capture because you got to do it quickly and someone has to have all that in their head because you look at this I need up transition from half guard

while this guy's pressing left and this guy like figuring out what all that even means let alone digging up the poses it's a lot of overhead Jeff often directed those shoots because he due to system warden I did and I was like you got to do this because I don't I was the only person who had everything in my head because I offered the state machine and it wasn't till way later in the project they're like our QA guy was up to speed and understood it Jeremy was up to speed and understood it was extremely complex and then the animation data like you guys are saying that the hand poses you get back from low-cap aren't good but if like the head is just resting on the shoulder and then it shifts a little bit that's better than having like a hand that has no no posing on it and is kind of floating and bouncing and stuff so there were some efficiencies there that you don't get from mocap and just you know usual it's with no caps and jittering some knock-knees some weirdness going to the ground yeah only this that's supposed to say but forgotten now thank you guys so much thanks you've got a huge combination of animations and various character sizes

you have a strategy for how you would test everything was working yeah it's really hard to do we would always test well I thought we always did but apparently we didn't the tallest versus shortest fat versus fat tall versus fat its Hall and fat versus short and skinny short and fat versus tall and skinny and then always men and women because there were differences with the female rig that it caused issues like breasts and stuff like that like colliding causing collision geometry to push you out in different ways we had to deal with some of that stuff too and but there's no way to test every combination is just impossible so just pick your extremes and like one middle of the road and hope for the best we didn't have any automated testing on that kind of stuff and the other thing I really wish we had time to do but we didn't was write like an automated script to walk the state machine basically so that a QA guy could just sit there and watch it do

everything and then like pick up on any problems and flag them but we ended up having like this master tracking document that was generated from the state machine that would have like a unique good for every transition and then one of our QA guys wrote a script that would pull telemetry from the game and like flag them as this has been played this hasn't been played so when QA was doing testing they could like flush that database and then retest and figure out hey no one's played this particular transition yet that's a problem someone has to trigger that to know that it's functioning properly but testing testing the system was a ridiculous undertaking to like every time something in the game changes you're supposed to do a full smoke of the system and this was just too large and complex to actually do manually so they're writing a lot about emotion tools we could have done a lot more but very very difficult to test and as far as like the animations testing there's a little trick that happens on pretty much every UFC name a title when the characters are in the takedown and they attained to the

ground if they scale at least bone length one two one the girth stays the same so if your mark on versus true mark how still fat Strube is still skinny but they're both essentially the same height and that kind of takers takes care a lot of that problem I think we're out of time now Jeremy and I will stick around up here at the front if anyone wants to ask questions thanks everyone for coming