Ubisoft and the Future of AI

The nucl.AI Conference – held in Vienna last summer – gathered together leading artificial intelligence programmers from various creative industries. Among them were Philip Dunstan, lead AI programmer at Massive Entertainment and senior AI programmer on The Division, and Chris Seddon, AI team lead from Ubisoft Toronto, who worked on Far Cry Primal. The two presented their AI designs at the event and spoke to how they overcame the challenges they faced along the way.

Philip and his teams faced many challenges, as they had to design AIs that would work effectively together and let The Division’s post-pandemic New York run smoothly. Meanwhile, Chris and the Far Cry team created an open world that feels alive thanks to dynamic wildlife, companion beasts that follow your commands, and a systemic design that ensures there’s always something happening in the world. Creating responsive and efficient artificial intelligence was essential to both open worlds.

Philip Dunstan, Lead AI programmer at Massive Entertainment

How would you define your mission as AI developers?

PD: I see my mission as all about supporting the player experience. A lot of my focus is working with the game designers and level designers to create an AI that matches the game’s gameplay and setting.

What do you see as the most interesting aspect of AI design?

PD: I love it when simple systems combine to create interesting behaviors. The complexity of our games nowadays has reached a point where it is extremely difficult to design and maintain a single system to handle all of the gameplay possibilities. Instead, we try to design smaller, separated systems to solve individual problems, such as deciding the best position a character should move to, or choosing which target an enemy should attack. When these systems are run together, the emergent result is often better than you could have achieved with a single design.

How did you overcome potential performance issues?

PD: The biggest impact came from decisions made early during the development process. The team knew that performance was going to be a challenge, and several key design decisions about the way AI-controlled characters move and how we create the player’s and enemies’ skills were very important to creating a good base for performance. I talked about both of these decisions at the nucl.AI Conference.
Later in development, it became very important to be able to measure the performance of the game in our nightly tests. This is a difficult task to do when you consider that we run servers [for The Division] with a thousand simultaneous players. To help measure performance, we created AI-controlled bots to run around and simulate the performance impact of that many players.

How did Snowdrop help you achieve The Division’s AI system?

PD: My favorite feature of Snowdrop is how it lets us iterate quickly on AI. Having spoken to a lot of AI developers at GDC and nucl.AI, I believe that our AI tools are some of the best in the industry. Our behavior editor lets us see visually how a decision is made and we can change the behavior of our AI characters while the game is running and see the results instantly. The Snowdrop engine makes it really easy for developers at all levels to create test levels to try out their changes. For AI, this was particularly useful, as we could quickly create test scenarios to see how the AI would respond in different situations.

In 2013, nucl.AI founder Alex J. Champandard predicted that the next step for game AI was scientific artificial intelligence. Do you believe academic AI and gaming AI are converging towards the same expertise?

PD: While there have been a couple of standout instances of academic AI feeding directly into game AI, mostly the two groups are trying to solve very different problems. It can be difficult to quantify the type of experience we want from the AI in the way that is required for academic research.

Nevertheless, it’s important that the gaming AI industry keeps an eye on what is happening in academic AI. Even though much of the research may not be that useful inside our games there is already a lot of evidence of non-gaming AI becoming useful in the tools that we use to create our games. It will be on the production side of game development that I believe we will see the biggest impact.

Chris Seddon, Team Lead Programmer at Ubisoft Toronto

How would you define your mission as AI developers?

CS: I think each AI developer on a team has a unique impact on a final product. For me, my mission is to create AI tools that give players the ability to create their own stories. With the right set of tools, each player experience can be different and exciting.

What do you see as the most interesting aspect of AI design ?

CS: The world of AI is exploding with new and innovative ideas. I think the most interesting part of AI design is the endless possibilities of what could be done to improve the player experience and immersion.

You experimented with Companion AI with Shangri-La’s tiger in Far Cry 4, which was a great success. How did it help you for Far Cry Primal’s beasts?

CS: The Shangri-La Tiger from Far Cry 4 was unique in that we had created the Companion AI to complement the overall Shangri-La experience. After having great success in Shangri-La, we wanted to bring Companion AI to the forefront of the action, so it felt like a natural stepping stone to migrate from linear side missions to the open world. Many of the basic “how do we handle that?” situations were known from our experiences in Shangri-La, so it allowed us to focus on making Companions that introduced new gameplay moments to the player.

Wildlife behavior needed to feel realistic while being viable gameplay-wise. How did you achieve that?

CS: Balancing gameplay and open-world systemic behavior that contributes to realism was very tough for us. Knowing that our companion’s primary role was to be used as a weapon, we focused on enhancing the gameplay elements surrounding the fantasy. To retain the realism of the animals, we allowed them to return to systemic behavior, when the player was idle. This retained the natural animal instincts the player would expect, while still maintaining their gameplay reactivity.

How did you manage to create a balance between the player’s enjoyment and a realistic eco-system?

CS: From the beginning, we wanted our Companions in Far Cry Primal to be fun and complement the player. When we had to choose between gameplay and realism, we chose gameplay. Fun is always more important than realism.

In 2013, nucl.AI founder Alex J. Champandard predicted that the next step for game AI was scientific artificial intelligence. Do you believe academic AI and gaming AI are converging towards the same expertise?

CS: I believe we’re entering a golden age for artificial intelligence where we’ll see some convergence of academia and gaming. I think it’s important for both worlds to work together while perusing their own individual objectives.