Understanding the AI design behind Ubisoft's biggest games

Designing artificial intelligence in games seems incredibly challenging, because your goal typically isn’t to build an actual autonomous intelligence; it’s to build an actor, something that can play along with a cast of characters to create the illusion of a real experience.

That’s well in line with the ambitions of Gabriel Robert and Julien Varnier, Ubisoft veterans who most recently led AI teams on two of the company’s flagship games: Assassin’s Creed IV: Black Flag and Far Cry 4, respectively.

“It's not about doing the ‘best AI’ -- it's about doing AI that have a purpose, from a gameplay point of view,’” Varnier told me last week.

“Yeah,” chimed in Robert. “Really, what we want to do is not build robots, but create actors that can understand the situation.”

The pair were speaking as part of a roundtable discussion with Gamasutra and other outlets at Ubisoft Montreal, where company employees from around the world gathered for the French firm’s seventh annual internal developer conference.

Building actors, rather than robots

They were also echoing sentiments commonly expressed by AI programmers to non-AI programmers, as when former Sims 4 programmer and current indie dev Bruce Hill told Gamasutra last year that “Most game AI is designed to be exactly as intelligent as it needs to be in order to give the player a good experience.”

But how, exactly, do you shape machine intelligence to ensure your players have a good experience? At Ubisoft (at least, the aspects of it represented by Robert and Varnier) the answer is to create opportunities for unscripted, player-authored events to happen; that means creating worlds filled with AI that can use and respond to gameplay systems like vehicles, alarms and elephants.

“Systemic worlds are really an important part of how Ubisoft creates engaging worlds,” said Robert. “We think that AI is key for this, and it needs to be shared with the organization.”

“Now everybody is trying to figure out ‘Okay how can we go more systemic, but also have the player at the center of the game?’” Varnier added. “So I really think that when we talk about this in our presentation, we are basically talking about creating an entertaining ecosystem.”

An entertaining ecosystem in Far Cry 4

Varnier and Robert were speaking to press after giving a talk to their compatriots on systemic AI design, the value of restricting access to subsets of your game’s codebase (only AI programmers can touch the AI code, for example) and what they’d each learned in their respective roles as AI tech lead on Far Cry and AI team lead on Assassin’s Creed.

The talk itself was highly technical; in that way (and in others) it was very similar to the sort of talks you see at GDC’s AI Summit. But in chatting about it afterwards, Robert and Varnier shed a bit of light on how and why Ubisoft designs AI systems for open-world games that may be of use to fellow developers.

“I think a sense of scope is really important,” said Robert, explaining that the business of AI design at Ubisoft -- and across much of the game industry -- has changed significantly in the past 15 years. Back then, he recalls AI as being quite simple, so it wasn’t always a massive amount of work to design.

“So often, from one production to another, you would see that people tend to redo the same thing because it’s quite fast, and you don’t really lose a lot,” said Robert. “Now that we have really big games with big scopes, we can’t afford this anymore.”

Robert recalls that when he joined Ubisoft, nearly a decade ago, he was “at the frontier of this, because I was always seeing people relearning how to do behavior trees, decision systems, perception systems” and the like.

Later, after the company’s inaugural UDC, Robert recalls developers at the studio being pushed to share more with other teams  

“We had the first UDC and we saw a big shift of strategy at Ubisoft to really go more with sharing, to be sure that we don’t reinvent the wheel each time,” he said. “Because as the scope and size increases, the cost of the wheel increases. So we really want to build on the shoulders of previous projects.”

Open-world game AI should put on a good show -- but not be optimal

Hence why, roughly half a decade later, Robert and Varnier are giving their own little GDC-esque talk to a room full of Ubisoft devs (with more watching remotely, via livestream or internally-shared recordings.) According to Robert, systemic AI design for open-world games is something the company as a whole has to focus on, and newly-minted coders don’t always get why it’s more important to make AI that does what the player thinks it should do, rather than what’s optimal.

“I think it can be a common mistake for junior programmers; often when they think of AI, they think ‘Okay, I want something smart, that can be optimal, that can really show that it’s found the right solution to the problem,’” Robert said. “People in AI programming love to solve problems.”

Developers on Assassin’s Creed and Far Cry games want their AI to solve problems, sure -- but only up to a point, and only in a way that makes the player feel satisfied.

“AI is not here to serve a solution; it’s here to entertain the player. To interact with the player. It shouldn’t look stupid (except if we want it to look stupid) but we don’t want it to be optimal,” said Robert. “Often, as programmers, we tend to try and solve problems instead of focusing more on good game design.”

Varnier agrees, and fondly portrays AI programmers as eternally caught somewhere between the technical and the creative sides of game development. 

“I like working on AI technology, but AI is also very close to design,” said Varnier.  It’s not about doing ‘the best AI’ -- it’s about doing AI that has a purpose, from a gameplay point of view. So we are really doing that tech conversation, but we are also part of the design conversation.”

Big budgets lead to big asset costs, which curtail designer ambition

That also means AI programmers are shackled by both the limits of technology and the tenants of good game design. At Ubisoft, they’re also restricted by the prohibitive expense of developing assets for a big-budget game.

“Sometimes we don’t do an actual AI feature or ability because the scale of producing all the assets to make it look like a 2016 game should could be expensive,” said Varnier. He suggested that from an AI programmer’s perspective, coding an AI to jump over a piece of cover is cheap, especially if it’s already designed to do things like climb ladders.

But the cost of actually animating that cover-clamber, then fleshing it out with some alternate animations (you don’t want your bad guys to move like identical robots, after all) and model-specific moves (slinging an assault rifle vs. one-handing a pistol, for example) might keep the move from being part of the antagonists’ repertoire.

“So obviously...at Ubisoft, and I’m pretty sure it’s true everywhere in the industry, that we’re trying to push more procedural animation,” says Varnier. “If the animation team can provide more animation with more flexibility, it makes the AI job way easier because you know your behaviors do not have a crazy cost in terms of assets.”

The same goes for sound design, as well. Writing, recording and mixing audio samples for NPCs costs time and money, and Varnier and Robert are hoping that one day procedural sound generation tech will get to the point where new AI behaviors can be designed and implemented without driving up asset costs.

“A lot of what make an AI look human, or sound human, is what they say. So it's a real AI problem, because being able to talk about what's going on in a systemic way, it's complex,” said Varnier.  

“And it's a big limitation, because every time we want the AI to say something smart, or say something about what's going on, we need to record those lines, we need to edit them, they need to be localized for five or six different languages. So again, it's still a very big upfront cost just to have the AI translate its behavior into something real, in the world.”

Here, again, Varnier and Robert are betting that procedural generation tech (in this case, vocal generators) will advance to the point where they can be used in Ubisoft's big games, which would loosen the big-budget bonds on the scope of what AI teams can try. However, both are quick to say that while AI designers love to get hype about new technology, they try to remember that neat tech and AI tricks aren't worth implementing unless they actually reinforce your game's core design goals.

"There are lots of things that are super-excting right now, but we need to be careful," Varnier tells me. "You need to think about, okay, what's the use? How do we build gameplay around this? Does it serve something?"

As an example, he points to the vaunted Nemesis system in Monolith Productions' 2014 hit Middle-earth: Shadow of Mordor. It's actually a super-simple system, he says, that doesn't involve any additional AI work at all -- but it makes the game's antagonists seem more alive, more real, just like good AI should.

"We can probably think about AI tech that probably makes that kind of design stronger, more flexible, and deeper. So there's definitely room for AI tech," adds Varnier. "But my point is more to say that obviously, we still need to keep that in sync with design and make sure that AI serves the player, in the end."