Maximising Impact with AI: Ethical Strategies for Developing a Top-Rated Game on a Modest Budget

Hello everybody. Uh thank you for coming to our talk uh titled Maximizing Impact with AI: Ethical Strategies for Developing a Top Rated Game on a Modest Budget. Uh if anyone wants to hear about unethical strategies, come and see me at the bar afterwards. Uh we're here representing VTime Games. Uh I'm Andrew Webb. Just a small amount of of my history. Uh I did a PhD at the University of Manchester in evolutionary computation. And then I was a research associate at that university in machine learning for a few years before I was the first machine learning hire at V time uh which now has a team of four um in the machine learning department which I lead. Um, I think probably most game developers don't have access to internal machine learning resources. So,

part of what I'm doing in this talk is justify my continued employment. Um, at this point I'm going to hand over uh for the first half of the talk to my colleague Jason. Thank you. Um, so usually at this point in the talk, we'll have a slide where I list all the games that I may have made. And I started to make that slide and it was uh a overwhelmingly Formula 1 games and uh b to too too many desperate things. So I actually just ended up putting the first and the last game on here just because it was it exposed an interesting couple of things. Firstly, the relationship between my first game, which is 1987 Cube to Kill, which no, nobody will have ever heard of, I'm sure, but uh that was for the BBC model B, and the most recent game, which we're going to talk about today, which is Borderbots VR. And obviously they share a little bit of heritage in the sense that uh the BBC model B begat the uh

Acorn Archimedes begat the ARM chip begat the qual the Qualcomm Snapdragon which begat the Quest hardware. Um but also just as interesting when I when I realized hang on a second they had the same price not adjusted for inflation. So 299 299 it's quite there's probably a whole talk just on that. Um but I'll leave other people who are more experienced with finances and uh economics to tell me why on earth even the the cost of the software is very similar. That's something that that sounds broken to me, but I'll leave that for somebody else to talk about. Um, to explain why we have uh machine learning at V time, uh, I'll I'll talk about the company very quickly. Um, we are set up by uh, Martin Kenright and we

we are a lot there are a lot of game developers in our in our studio. We're relatively small, but we still thought it was good to invest in uh machine learning. Um we do we do do games. We are from a gaming background, but we also do gaming adjacent things. Lots of immersive stuff, lots of uh what we think are innovative uh application development. And uh yeah, we're all there up in Liverpool. Like I say, we've got a smallish team, but we still think it's worth investing in machine learning. Um, the most recent game that I'm going to be talking about is uh Borderbots VR. There's all sorts of nice uh marketing quotes there about it. Um, and I'm going to talk for once. I'm not going to do the technical talk. I am the CTO of the company, but I'm not going to do the talk about the technology specifically. I'll I'll do the story bit and I'll leave the the tech bit to Andrew. Um, so this game came from a a

relationship that we had from quite a way back. So I'll just quickly explain that. Um, in about 2018, a company called Pawprint Games made a vertical slice of this game called uh that which they called at the time Border Patrol. And they tried to get somebody to publish it and at the time no one was biting. um VR was still in its infancy. The Qu It was prequest really. And uh even the fact that that the a YouTuber picked it up and made a a great little uh introduction to it um and really sold the idea. Um they didn't get any uh backers at the time. I think that the the the easiest way for me to explain the game is just to play a very small segment of this uh YouTuber's um interaction with the game. Is it going to play? Can you press space

on there?
Listen, man. I don't want to be the futuristic TSA or anything like that, but I think I need to poke your glowing robot nipple. I'm just going to Oh, okay. That's a whole lot of virtual reality smuggling cupcakes. shoot this man. [Music]
So that got the idea over I think really you know in 10 seconds. That's a brilliant explanation of it. Um it's particularly to pick up okay it's it's it was called the the video was called um papers please in VR. That's obviously the high level concept. Um but it also got over the idea um and there was a little bit of um reticence because it was uh you know a border patrol a border thing has a few connotations that may not be um positive uh in the current climate but very that proved that if if you just put add humor to it it it you

know you can you can get past that very easily. So, uh, Pawprint didn't sign that game and quite a few of the employees ended up at V time. U, we had a a few of the core members of it and Team 17 a few years later looking around and looking at the space said, "Hey, did you ever make that game?" And the answer was no. And they said, "Well, actually, we we still think it's a great idea." So, a little bit of no gayiation. And then because we had the advantage of quite a few of the members working at our place, we got to make the game and we worked with them to make the game. Uh there was a little bit of a problem that all we needed to we had a vertical slice. Great. All we needed to do was add a story of which we hadn't had a lot of experience with. So we and we brought in international hobo to do script writing. Uh and then we just needed to just needed to put together the narrative and we had a lot of help here. Jed Ashford from uh

realized realities helped us brainstorm all that through. Um so we felt then that we had enough knowledge. Uh and Steve Napman there in the in the uh in the uh picture there he's the main designer. He had a very strong vision for what the narrative how the narrative should work and how it should be different from paper papers please. And of course papers please quite dark. We had to go completely the opposite direction and that took us into an unknown uh you know that had unknown consequences at the time. Um my my clicker not working now that yeah thank you. Um, so again, we're quite new to this. We looked at the tool sets that were available. Um, Arweave, the designers loved Arweave. Uh, visual scripting stroke bolt from Unity. Coders were like, "Oo, yeah, this is the thing." Um, this is Twine, which is

another narrative uh description tool. Uh, couldn't use that light background, not not game ready. Um and finally articy draft which is a quite a complicated and complete toolkit for writing um interactive narrative. Um and we got quite close to adopting that. We did quite a lot of due diligence on in the end. We decided it would probably be a leaky abstraction between RTC draft and bolt which we would be building quite a lot of the stuff in and that the interface between them would get messy. Um so we ended up building almost all of it in bolt or at least that was the intention. At the same time we have all these tools that are inhouse and we have now bear with me we've got two different conventions for naming our code names. Uh I call everything uh from a code point of view we're named after people. So Awson is our texttospech uh

model uh just the code name. It's not a product name. And then the project manager would like to name things after mountains. Um so we had awesome a texttospech model. Then we had hvellin which is a way of giving an actor or a person a script and asking them to to basically uh read out the script uh with prompts for style changes and for us to gather all that audio which is then properly annotated. So we had that lying around that was allow us to make uh voice cloning. Um as I say we do lots of other things beyond games. So we had some of this technology lying around. And then finally Vtag which is an avatar based uh messaging app. And that had a load of tech that we in it that we had lying around as it were for recording face animations and putting them on avatars. Um so we thought oh some of this some of this might be of use in the app in the game development.

Once we started pro uh pre-production we realized that actually these nodebased uh solutions were not actually that useful for us. Um even at that st at that stage we decided oh we want it to be kind of funny. So the designers were trying to build stuff in uh with with with um nodebased tools but of course most of the comedy is in timing and things like that. So the timeline became very much more important to anything else. So actually we're probably as well to have avoided some of the nodebased tool tooling although it did mean we had to write more stuff ourselves. There's one thing specifically missing from here and which would have been fantastic and if Unity uh somebody from Unity timeline team could tell us a way of getting waveforms on there without using their waveform class because this

is our own clips which are doing various animation things without the waveforms on you can't see where the actor takes a breath or what have you. It was really annoying. Um, I found a way in a in another project to get them on there, but it's insanely expensive and probably wouldn't have survived production use. And then finally, once the script started to arrive from International Hobo, we realized several things. Firstly, it was funny. It's brilliant. Uh, so we had lots of ways of trying to get the script to work as it were. Obviously, the timeline's great. We can we can queue up the jokes. We can get the responses at the time. Awesome probably wasn't the right thing. Robotic voices because in our heads we're doing a game about robots. So, we we're like, "Oh, well, it won't matter if the voice is a little bit robotic." Of course, it does. It

does. It would if it was a different type of tone perhaps it wouldn't matter so much. But because it's trying to be funny, it needed a human. So, we did some am stuff. So, this is the back end of Hellen. We put all the scripts in hell and we did some amrram stuff just to battle test it and unfortunately that video could you click that video sorry again. Yeah we in along with Helen we added face capture just you know just to prototype it. And finally, um, we did some tools where we could just grab a set of lines and then say, okay, make make a big make me a timeline out of that and then it would just be a matter of, um, arranging everything uh, for it. And that gave us enough um, experience and it gave Team 17 uh, enough uh, of a an impetus for us to go to the studio. We' we fully prototyped it. We had the entire script

in there. We'd edited the script quite heavily and apologies to International Hobo. We took an absolute chainsaw to their script in in parts. They'd done a very Terry Prattet quite British uh word playheavy script, which worked in some parts, but because you're in VR, time in VR is really precious. It's not the most comfortable place to be all the time. and listening to extended tracks of uh very funny dialogue isn't always the most uh comfortable thing that you want to do. So, we'd hacked it all down and we we had very good um we had very good idea of exactly what the script would be and we could go to the studio and so we got proper actors and uh this is a VR title. So, this is quite high up on the uh relative budget terms for uh a VR title. you know, a lot there's a lot of titles

out there that are just one or two person teams. We're we're a although we're still small by gaming standards, we're we're quite big by VR standards. And we put the effort in. We we got uh actors in and as you can see the the experiment was can you just plon an iPhone there? Now, we were very lucky because there was a pandemic and the reason that we were lucky for the pandemic is the software was a bit clunky and it forced the actors to actually find the line. It was in it was in numerical order, but um they had to basically hit record and play before and after every clip. We were we we group things together. So, it wasn't every single line, but it was between every segment. And because of the pandemic, they had done a lot of home recording. They'd been stuck at home and they'd had to get their own audio rigs up. They'd done their own uh

soundproofing. And for them to work, they basically got all these skills together. And so we this was quite a machine. So the the learnings from the machine rather than machine learnings are just uh this arrangement where international hobo were in on the call. I was there international hobo obviously sort of giving uh exposition and telling them what the the the the scene or the the dialogue was meant to to convey. I was there quickly checking our graphs and seeing where we could do cuts and where we couldn't do cuts because they would obviously the actors would like to run multiple sentences together and do uh large chunks together. So I was saying where the chunks were. The actors were coming in, the director was firing out instructions, choosing takes, and then the engineer, amazing, was hitting record while he was chopping up the previous takes sentences.

And the Helen was there recording both the audio and the video of the facial capture from that actor which proved to be very important that we had the audio because when we got it back we were intending we thought oh we'll just record timestamps and we had timestamps between the high-end top grade audio that we got from the studio and we thought well we'll just match that with when the recording for the vid for the facial capture happened that didn't work in the first couple of sessions in we didn't have any time stamps at all because of a bug we tried putting the time stamps in and then we couldn't get them accurate enough so that we would match the video uh or the facial capture so in the end we had to well I I held my hands up and uh went to Andrew and said help I've got this problem of course machine learning guys go ah ah that's just easy so they gave us a Shazam like

algorithm where I could give them the high-end audio, 24-bit audio and our 16 bit audio that we got out the phone and it could match the exact point at which they they they matched um and then we could match the implicitly then ma uh match the facial animation and we also had to write a load of tools because as you saw on the previous I'll just go back to the there you can see the the the phone is off to one side and of course it's not going to be Perfect. And it's also especially one or two of the actors like to read the line. The line, you can see them reading the lines there. Some of them didn't like to read the line or look at the the script while they were doing their performance and would turn and they turn away from the camera. So often, well, one of the actors in particular, their their eye would just do really weird things and would just be spiraling away in the corner. So we had to come up with little tools where we could preview all this stuff and through the magic of the

machine learning that Andrew is going to talk about in a minute we could infill all these parts. So we managed to uh get most of these done and in a few cases of course we didn't get a match. We didn't know we didn't have a recording or we couldn't find a recording that matched a piece of audio. uh that we've got possibly because of an error in recording. Usually it's an error in matching that could that the algorithm couldn't match it. So in that in those cases we could either generate the whole animation and that would work again another 90% of the time but where it couldn't we actually you see the QR code coming up there we would actually re-record it at half space half speed. And actually Jolie, who did the uh some of the animation recording, she was really good at mimming along to the record the the recording from the uh actual actor and then basically miming

and we would record her va her facial expressions and play them over the final thing. So we got it all done and we got it all done pretty cheap. This is VR project. We can't spend a lot of money. We can't spend hours with animators and things like that. So, the fact that we had so many backup strategies, we would record the face. If we can't record, if the face isn't quite right, we can tweak it. If we can't tweak it, we can generate it. And if we can't generate it, we can reddo it using our recording stuff. The main thing that I was amazed at was how good that iPhone was. Um, I went down expecting that I would because it's just pre-production software running probably running in debug mode and all what have you. I thought that it would overheat. I would have to be recycling phones, all this stuff. No, one phone was enough. I could have I could have just had one phone and just charged it at lunchtime and it would have got through. I did. Obviously, I had backups for those. Um,

and then finally, the QR codes. Obviously, again, it's another pandemic learning. Everybody was understanding, okay, I could give the engineer an A4 with the call sheets of all the actors who are going to be doing their stuff that day. And he in the end, although I did the first week of recording, I just handed it over to the uh engineer and just said, "Here's here's an A4 piece of paper with all the actors. Just scan the QR code and load it onto the phone." Cool. That's how we got it all done. Now to tell you the actual nitty-gritty of it, I'll hand you good back over to Andrew.
Thank you. Uh yes, so I'm going to I'm going to give some details about the use of machine learning in in the development of Borderbots VR. Uh I'll be giving three examples. Um we used a mixture of uh so some of the tools that we used were publicly available tools available to anybody. Some of them were in-house tools um that were the result of

extensive uh research and time and effort. Um I'm going to talk a bit about the advantages of having that in-house expertise. And uh you might ask why ethical in this title. Uh really it's the so the focus for us in using machine machine learning in development was to eliminate repetitive work and drudgery to enable the team to do things that they otherwise wouldn't have done but not at any point to replace creative work. Um so for example there's no there's no generative AI in this talk. Um, so example number one, uh, this was the assignment of animations to our non-human characters. Uh, so the game is border bots. A big chunk of the characters in this game are robots. That means they don't have expressive human faces. So the the way that they emote is with whole body animations. Uh, so to see some examples, this character is

he's a toaster. His name is Bernie. Uh, and he's he's very happy right now. Uh, now he's sad. I think a bit less sad. I'm not sure about that one. Uh, and now he's terrified. Um, and on the left hand side of this slide, we have uh some of the script lines from the games from the game. Uh, so this is the script as it came back from international hobo. We've got a line up here, up here, and then associated with it is a direction which is intended to be read by the voice actor to give them some direction for, you know, emotion, inttonation. another line bleep bleep bleep to be done in a singing way. Uh you I'm not a voice actor so I don't know how to do that. Um so the problem is how do you assign the appropriate we want an animation for each line of the script and it's an easy job to do. You know what's the appropriate animation for this line? It's the scared one. It's easy. But we had uh 1,350 script lines that needed animations assigning in this way. Very

tedious to do manually. very timeconuming. So this is an example where we used a tool available to everybody. We used uh the enormous hammer that is GPT because this looked like a suitable nail. So uh probably everyone in this room's used GPT probably almost everybody has chat GPT the web interface. Um, we were using the GPT API, which really I think, you know, for me it's less of a chatbot and it's more of a text completion service that excels at performing tasks that it's been given a very small number of examples um of. So what we do here is for every line of the script, so here it's the blue line that we care about, the bleep bloop one, we first give it three other lines of the script, always the same three examples. Up here, up here, panicked. And then on the right hand side of the arrow, the name of the emotion that we want and the intensity that we want. Three examples that we've we've come up with. Uh one of them with

a missing direction in case any of them are missing. and then the line that we actually care about. Uh give it to GPT and get back happy intensity six. Um this is this is quite a powerful and large tool to use for this problem. So you might think why not use a simpler approach? Well, it works and it's flexible and we had experience using the API and it's cheap. Uh this took I think two hours to put together on our end including you know come up with the prompt, mess around with it, tweak, test, make a little service, hand that over to the devs and maybe a little bit more on their time on their end to integrate it into the workflow. But yeah, very straightforward and saved tons of time. Of course, it can make mistakes. So some of the animations will be assigned incorrectly, but then that gets caught in QA or play testing. We manually change the animations.

Uh let's see a scene from Borderbots VR with uh this in action. So all of the characters in this scene are robots.
Hey.
Oh,
do you know how difficult it was for us to set that up?
You promised you wouldn't say I helped.
If I'm going down, I'm taking you with me.
I deserve it. I am guilty. I helped Bernie recover some of the old toasted slices out of the bins. Send me to jail.
What is wrong with you?
What's wrong with me? What's wrong with you? You stop us doing anything fun.
I stop you burning the apartment down.
Exactly. I might produce perfectly shaped toast, but you're the real square here.
That was good. You should write a screenplay. A square, am I?
Yeah, a complete stickler for the rules.

I bet you don't even overclock your processors.
Would a square do this?
Oh, my circuits. Do you mind? I'm not properly earthed.
So, uh, incredibly dramatic scenes. I think you'll agree. Um so yeah all of the animations there chosen by this um little language model tool. Uh the next example I'd like to talk about uh this is uh so the previous example was using this tool that's available to everybody. Um for this example we used awesome which as Jason mentioned earlier is our in-house texttospech engine. Uh you might ask why do we have our own texttospech engine? Uh services exist. I'm sure there are people in this room who subscribe to highquality texttospech services. The first answer is that when we started working on ason there were no such high

quality texttospech uh services. Um and now that we have it why not use it? Uh the second answer to that question is having this inhouse means we can have exactly the features that we want. Um, so for example, early in the development of Borderbots, uh, we thought that we might need for our text to speech, uh, English, French, Italian, German, and Spanish. So, we added support for those languages um, relatively quickly. I'll talk uh, a bit more later on about another custom feature we were able to have because this was in-house. Um, let's uh, hear this in action. Ah, good morning. So, they finally sent over that missing toasting unit I requested when we first moved in. Yet, we didn't get quite what I was expecting. You should take a look. It is still in the delivery drop off area. Hey, I'm Bernie. Pleased to meet you.

I'm here for all your toasting needs. I can't wait to get these grills going. How about we get toasting? I need to calibrate my burn. Grab me some bread from the bread bin over there. I can't wait to use this grill.
So, uh, notice that wasn't the same audio as in the previous slide. Um, the reason is quality. So, text to speech has come along quite a lot in the last few years. It's come along uh since this, but it's still not either in audio quality or in the quality of the voice, the sort of emotion of the voice, the the intonation, the human touch. is still not quite there. And as I indicated earlier, we're not really in the business of trying to put creative people out of work. So, our use for text to speech in this game was purely as placeholder. Um, so it meant as soon as we had a script from International Hobo like that, we had all of our audio in place

as placeholder, over three hours of audio. Um, and then whenever we got a script update from IHOBO, we just regenerate those lines. They're immediately updated. Um, the clip that you just saw also had subtitles going. Uh, so one of the other things that we get from Awson is uh timings for the start of each word so that we can align subtitles. Just another feature that we wanted so that we could include it. Uh, why bother including placeholder audio? you know, we could have just had silence and subtitles until we had the voice actors in in the uh recording studio. It was really helpful for us actually to have placeholder audio, particularly in editing, um getting timelines right, knowing how roughly how long a certain line's going to take so that you can build the scene around it. If you want an animation to be triggered for emphasis on a particular word, it's helpful to know roughly where that word's going to lie. Um, something Jason

alluded to earlier, uh, when you're reading a script off a page, it it's not always clear, oh, this line is far too long, but if you're in VR listening to it, you're like, I'm I'm totally bored at this point. Um, there were other things that were so more generally we were able to send clips of gameplay footage back to International Hobo with our feedback and drive iteration and script changes through being able to hear it. Um, as Jason said, it's supposed to be a very funny game. Um, and even with even with text to speech, which obviously has no sense of comedic timing, it's much clearer hearing a joke, whether it lands or not, than reading it off the page. Uh, other examples, we had a character whose gender changed as a result of this kind of feedback. Uh, Bernie, the toaster, who you met earlier, in the initial script, Bernie had a lot a lot of innuendo. uh and on on play testing we just

decided this isn't working at all. So that that all went um so lots of feedback lots of iteration as a result of this. Um and and the other key point here is we didn't develop Orson we didn't develop our text to speech for border bots. We had it as a result of previous um previous work. It's lying around. It's very cheap for us to use. uh we worked out using Autton is about hundred times cheaper for us than using one of the leading texttospech services. So why not use it? Um and as I said, we're not replacing voice actors. And I think you'll all agree when we uh hear it here, this is this is this is what you want in a game. Um especially a game with humor and character.
Ah, good morning. So, they finally sent over that missing toasting unit I requested when we first moved in, but we didn't get quite what I

was expecting. You should take a look. It's in the delivery drop off area.
Hey, I'm Bernie. Slice to meet you. I'm here for all your toasting needs. I can't wait to get these slots fired up. How about we get toasted? I'll need to confirm my calibration will produce a perfect golden slice. Grab me some bread from the bread bin over there.
Uh so, uh if you were paying attention, you might have noticed there were some script changes there. So, same scene, uh slightly different lines. Um partly due to this, yeah, iteration we were able to do. And the last example I'd like to talk about, this is another case of something that was inhouse. Um so this is you know uh a model that we designed we developed ourselves we trained it ourselves with data we collected ourselves you know it's all all in-house and this is our speechtoface animations model this is

for the human characters in game and the purpose of this model is given some speech audio for them uh drive the face appropriately. Uh so the way it works um there are actually two models. There's one that from the audio uh determines what the character is actually saying at each moment. So which part of speech are they saying? And then there's another model that takes the output of that plus the audio again to drive the face appropriately. Um it turns out the mouth is much easier to handle than the rest of the face. If you know what somebody is saying and how loudly they're saying it, you can pretty convincingly drive a mouth. um to sync with that. The rest of the face is harder just because, for example, from what I'm saying now, if you have your eyes closed, you don't know if I'm waggling my eyebrows or, you know, wobbling my head around. You don't know. So, um that's a little bit more complicated. Um I said here we gathered the data ourselves as well. Um it's worth saying something here about the flexibility of the Helen tool that Jason

was talking about earlier. So because that tool um what it does is records speech and face um of a person reading a given line, it means that we use it to gather training data for this model. We use it to gather training data for the texttospech model and we use it to gather face data for the final game from our voice actors. So very versatile tool. Um, as with the text to speech, if we had any script changes, uh, we could just rerun this model over them, get new updated face animations. And as with the texttospech model, we didn't develop this for Borderbots VR. We already had this, and you know, we're reusing a lot of technology. We were using the same face parameterization that we' used in previous projects, so we could just drop it straight in. Uh we'll see an example now. This is Pandemia, uh a human character in

Borderbox VR. She's the CEO. Uh and we're going to see this is a clip from early in development. Uh so we'll have both the voice and the face being driven by ML. [Music]
This is Pandemia, the CEO of Fortitude. You may recognize me from TV or the many statues. Oh, I think that's an older video from a previous version. Okay, it was supposed to be a longer video, but never mind. You get the idea.
Um, and then for completeness, let's see. Uh, this is the same character. This time it's not the same scene. Um, later in development, so now the voice and the face are both um coming from the voice actor. The face coming from the iPhone recording. Citizens of Salvation Heights, this greatest of cities, we find ourselves in unprecedented times. You may have noticed weward behavior from some of the robot workforce, the gears driving our

fine city. Well, to ensure maintenance is being completed, and for everyone's safety, a series of checkpoints have been established around the city to manage road. I'll just interject there that there are a few bits of ML in there as well. Um you know there are this is one of the actors who would look away and so the CEO particularly had roving eye syndrome. So there there will have been eyes erased and replaced with ML at points.
Yep. So yeah um putting it all together this was the situation early on in development for our robot characters. We had an emotion selector that would choose how to animate the character and we'd have a text to speech model that would drive the speech. So we had straight away as soon as we had the script we had both speech and animations. Later on in development for those robot characters we had the professional voice actors very very good

performances but we still used the emotion selector. So that that was used throughout. And then early on in development for our human characters, we have a texttospech model giving us the speech and the output of that is fed into our speech to face model which drives the face. So everything is ML. And then um for our human characters later in development, everything comes from a professional voice actor ideally, but due to the issues that uh Jason alluded to wherever there are problems, we can drop in either our own re-recordings of uh facial expressions or we fall back on the speechtoace model. So that's still used uh throughout. So what lessons can we take from the use of machine learning in game development? uh you must understand and embrace the possibilities. Uh machine learning is a very fastmoving field more so than tech generally. Uh

all of the examples I've talked about today are possible due to advances made in the last four years. Um that means you need to keep on top of things. If you want to know what's possible now, what's possible in six months, what's possible in a year and those are going to be three very different answers. you might need an ML hire. Um, as I hope have indicated, there are advantages to having in-house expertise. Uh, this allows you to do some things that you otherwise wouldn't be able to do or in some cases, like in the text to speech example, it allows you to do things that other people will be able to do, but maybe a year earlier, maybe two years earlier. Um, and you won't just be able to press a button. So if you are um working on things in house, this is that this is going to take a significant investment of research, time, iteration. Things don't always work out first time. You don't always know how uh things are going to go. And at that point, I'd like to say uh

thank you for attending the talk and we'll leave you with a trailer for Borderbots VR. Uh so enjoy. We have the audio up. [Music] One. One. [Music]

[Music] Some birds people in your dreams. [Music] Thank you. [Applause] How we doing? I think we maybe have time for a question.
Yep. Uh just wait for the microphone.
Cool presentation. I did. Yeah. Really cool presentation. Lots of sort of cool tech. You introduced that. I guess I was just kind of curious and now thinking about it, maybe it wasn't sort of your decision, but there's a lot of tech that almost seems sort of underutilized on the sort of robot side of things. I

guess I'm just sort of curious if there sort of future plans to introduce that into more of a human sort of product versus a sort of robot robot.
Yes. So, you're talking about the fact that the facial animation didn't make it to the robots. We only used it with the Yeah. So obviously the facial animation side of things was exper very very experimental and we didn't know it was going to work quite so well and probably we would have you know like to have uh exploited it more. Um and in fact we actually built our own libraries that was one of the takeaways that we did did learn from ourselves was oh we need to have used that. So one of our um immediate tech demos afterwards was hey could we have applied that to say a couch. So one of the artists made a couch with all the bits on it and we built an entire library uh so that we can very easily apply all this stuff to anything not just faces.

Uh any more questions.
Okay. Thank you very much. Cheers. [Applause] [Music]