Read this article in French and many more from Ubisoft Stories.
Discover the job and career of Rémi Poivet, Cognitive Science Researcher at Ubisoft and author of an ongoing thesis on the “credibility” of NPCs in video games.
It may come as a surprise to some, but researchers do work in different Ubisoft Studios, including research and development departments. Rémi Poivet, 25, is one of them. For Stories, he talks about his career path and the reason why he is at Ubisoft Paris: to understand non-playable characters better and help make them even more believable.
What is your connection to video games, and how did your academic background gradually lead you to Ubisoft?
I’ve been playing video games since I was very young. I think I started on the GameBoy because I remember Tintin and the Temple of the Sun. I still play today all types of games when I have time, parallel with my Ph.D. However, I had never imagined myself working in video games. Being a big fan of movies, I first wanted to do an Audiovisual BTS (Higher National Certificate) with the view of directing. But I did not regret finally doing a Bachelor’s degree in Psychology, where I learned a lot and was particularly interested in studying the brain. So I did a Master’s degree in Neuroscience in Tours. When I embarked on my thesis, I wanted a subject at the crossroads of cognitive science and my desire to study artificial intelligence (AI), which had become my specialty. I turned to “autonomous agents,” physical or virtual machines controlled by computers with the objective of human-machine interaction. Hence the importance of studying perception to create better autonomous agents.
My thesis supervisors mentioned several possibilities, but I finally wanted to submit a more applied Ph.D. with a CIFRE, i.e., a collaboration between a company and a research laboratory, and then apply at Ubisoft. I thought it would be great to do my thesis in a company like this. After getting in touch with someone from La Forge in Canada, the head of R&D in France contacted me again to join Ubisoft Paris. I am now part of the Data Science team. In parallel with my thesis, I work a lot with the production teams at Paris Studio and the User Research Lab, which organizes the tests with the players.
Why are researchers recruited by a company like Ubisoft? It may seem surprising at first.
Evidently, video game people have a genuine interest to always go towards innovation; they ask themselves the same questions as animation studios such as Disney and Pixar in their time. Moreover, innovation is intrinsically linked to the world of research and development. So there is a natural desire at Ubisoft’s Studios to push this R&D, and I find it very interesting that the company relies on the work of researchers to achieve this. It is undoubtedly becoming an actor in the world of research, which is developing more and more at the industrial level. Since Ubisoft’s ultimate goal is to improve the gamer experience, it was an obvious choice for them to get involved in this field.
How did you get involved with the artificial intelligence of non-player characters, and what is your goal with your thesis?
I participated in a study that looked at our ability to attribute human characteristics to the famous virtual agents I mentioned earlier. The work was on the frontier between philosophy of mind, AI, and psychology, and I truly liked that approach. I also had the opportunity to take courses in AI and do an internship in the United States in a laboratory specialized in the subject. It was my first step in the field, where I learned basic programming and the different techniques of AI. Then, when I embarked on my thesis, I wanted this work to join both AI and cognitive sciences. In the end, my subject is more about “non-player characters” than their artificial intelligence. The AI of NPCs is the decision trees, how they behave, move, act in their environment. The NPC is the AI, but also their appearance, the perception we have as players, and the cognitive biases that can result from it.
Can it be explained why some NPCs are perceived as not being very intelligent or not challenging by some players? Can we say that it is possible to make truly intelligent NPCs?
There are, obviously, bugs, but how NPCs behave is also fundamental in perception. In the early 2000s, when decision trees were first being integrated, studies showed that an enemy was perceived as intelligent based on how difficult it was to be defeated. Intelligence was a challenge. Now, and especially in open-worlds, an NPC can not “just” be difficult to beat to be perceived as intelligent. The subject has a real complexity, especially since the games are not all housed in the same brand; some AIs will be better “perceived” than others by players. It is called the halo effect. Typically, many people like the NPCs from Breath of the Wild; however, they have a relatively simple design. Now, we must consider the notion of believability, credibility, and our ability to believe in a non-player character.
Can you explain what exactly believability is and how central it is to the implementation of NPCs?
The concept of believability was already evident through Disney in their early days when they sought to make their characters more charismatic, not realistic but believable, integrated into the universe in which they evolve. Lasseter and Pixar continued this work. Today in our games, when we say that a character must be believable, we talk about their ability to integrate into their environment and react to their surroundings.
Many studies have been carried out on this subject. They discuss the importance of personality, what makes a character unique, and his ability to interact with other agents and the environment or his coherence with our internal knowledge as an observer.
How can we succeed in measuring this believability which, by definition, is very much a matter of subjectivity/perception?
In the scientific literature, there are scales of measurement to judge this believability. But generally, these are scales broad enough to be applied to very different situations. As part of my thesis, we are still reflecting on our own scale in the continuity of what has been proposed until now. The idea is to determine the parameters on which we can act to adjust the believability. I have already been able to work on the consistency between the appearance of an NPC (his clothes and equipment, for example) and his behavior (hostile, friendly, etc.), and I will continue to test other parameters.
Written by Vincent Manilève on Stories!
What processes or tests do you rely on? How do you make this notion concrete for players who are sometimes novices and do not always have the correct codes?
We are working on hypothesis verification in a very controlled context. In a first test, we will interact with a single NPC whose appearance, behavior, and position of the player we will control. It is a different framework from a classic game. Still, once we have verified our hypothesis, we will test it in a regular game context, for example, during playtests with the User Research Lab.
As for the players, it is necessary to select the participants according to our research hypotheses, and therefore, sometimes we have participants who know the game, sometimes not. It is essential information to consider when studying perception. We also plan to create a detailed questionnaire on the subject, aimed at them and which we hope to disseminate widely.
Ultimately, how do you create a good NPC that is more rewarding for players and their gaming experience?
Creating more credible NPCs is my thesis goal. It is necessary to improve their coherence, their internal references so that their behaviors are perceived as rational. When we know what a Rabbid is, we have clear expectations. Still, we also expect it to be rational in its irrationality. Essentially, the Rabbid must be a moron. It is also how I define intelligence in the context of my thesis: we must have a rational behavior with respect to our internal references as an observer. That is what we are working on at the studio. We hope it can help production teams develop and deliver more believable and consistent non-player characters than ever before.