Arc Raiders is one of the biggest breakout hits in years, and today's new Flashpoint update promises to make it even better. And it was all done on a smaller budget than you might initially think, compared to the biggest blockbusters on the market.

And it doesn't take a trained eye to see that, even at a glance, the enemies in Arc Raiders feel fundamentally different from traditional game AI. They don’t follow rigid patterns or scripted behaviors, but instead, they react dynamically to the environment, recover from disruption, and occasionally end up in places even the developers didn’t anticipate. That sense of unpredictability is not just a design choice but the result of years of research into robotics, physics simulation, and machine learning.

At Embark Studios, the team approached enemy design from a systems-first perspective, treating enemies less like animated characters and more like physical entities that must navigate and survive in a dynamic world. That decision led them directly into robotics research and reinforcement learning, borrowing techniques for controlling real-world machines and adapting them to a game environment.

Rather than relying purely on traditional AI systems, Arc Raiders blends learned locomotion with behavior trees, creating a layered approach where movement itself becomes part of the intelligence.

We spoke with Martin Singh-Blom, Machine Learning Research Lead at Embark Studios, about how these systems work, the challenges of bringing them into production, and why physics-driven AI fundamentally changes how players interact with enemies. Below is a lightly edited transcript of that in-person interview.

Can you introduce yourself and summarize what the focus of your talk was at GDC?

Martin Singh-Blom: My name is Martin Singh-Blom. I joined around January or February of 2019, so just a couple of months after the studio started.

The general gist of the talk is that physics is a very important part of Arc Raiders. When you shoot drones or enemies, they react correctly, and a lot of the gameplay comes from emergent interactions. For example, you might shoot an enemy into the wall, and it takes extra damage and explodes from flying into the wall rather than from your shotgun.

We knew very early on that we wanted emergent gameplay like that because it leads to high replayability. We also knew we wanted legged enemies, big robots that walk around and interact physically with the world.

If you want to have that, you run into the same problems as robotics. Even though your robot is simulated, it’s still a physics robot that needs to move and balance. So we had to go into the robotics literature and take their methods for controlling actual robots, put that into our game, and make it work for us.

A lot of the talk was about what problems we encountered, where it’s the same as robotics, and where it’s different. For example, robots can use cameras, but we can’t do that because it would require rendering on the server, which is too expensive. At the same time, we can cheat—real robots can’t have magic forces that lift them up when they start falling, but we can.

And then a big part of it is reinforcement learning, which is a deep learning method people use for robotics. So part of the talk was introducing that, and part of it was about the challenges of applying it in a game.”

What makes Arc Raiders’ enemies feel so different compared to traditional AI?

Martin Singh-Blom: I think not all of this comes directly from the robotics work, but almost all of it flows from the decision to make everything physical.

In a traditional animation system, you always know where the enemy is. There’s no flexibility. If you bump into it, it doesn’t react because it’s just playing an animation. But since we decided to go with physics, we don’t necessarily know where the enemy will be. If you throw something at it, it might get bumped out of its path and end up somewhere else.

That forces us to design all the systems to be much more flexible. A lot of the strict pathfinding you might have in other games doesn’t work anymore.

So what players perceive as intelligence is partly that the enemies are acting in a real world with the same degrees of freedom as real objects, and partly that they are designed from the start to correct themselves and adjust in that environment.

Then the rest of it is honestly less magical than it seems. We have a really strong AI team using things like utility AI, and they’ve spent a lot of time tuning behaviors to feel consistent and punishing.

We don’t design enemies around patterns. Instead, we think about what principles guide the enemy and why. That makes them feel like they have intent. It shifts the game away from memorizing patterns and toward understanding behavior.

Why doesn’t the game rely on traditional enemy patterns or scripted encounters?

Martin Singh-Blom: We can’t really do that because of the physics.

We’ve tried building more traditional behavior systems, but as soon as something unexpected happens, like an enemy getting pushed, it breaks those patterns.

It all comes back to physics. The physics adds a layer of unpredictability that seeps into the design's fundamentals. So we had to move away from that approach entirely.

How does machine learning fit into the overall AI system?

Martin Singh-Blom: The machine learning part is actually more limited than people think. It’s strictly for locomotion, such as how the robot places its feet and moves. That’s a very hard problem, and traditional methods don’t work well for legged robots, so we had to go to the research frontier and use reinforcement learning.

For drones, we don’t need that since we can use traditional control systems, as you’d see in real-world drones. But for legged robots, we haven’t found any other way that works.

Once you move into higher-level decisions, like where to go or what to do, that’s handled by more traditional systems like behavior trees.

There’s a boundary between the two. For example, if there’s a box in front of the robot, the behavior tree might decide that it wants to move forward, but the locomotion system decides how to get there, whether to go over the box or around it.

As the models improve, we can push more decision-making into the machine learning side. That’s where it gets interesting, because the robot can start making its own decisions, like deciding to squeeze through a space or jump over something, and that creates more surprising situations for players.”

Can you talk about some examples of emergent behavior and unexpected outcomes you've seen?

Martin Singh-Blom: We’ve seen situations where enemies end up in places we didn’t expect at all.

There was a video internally where someone showed a leaper inside a dome, and we didn’t even know it could get in there. We have no idea how it got there. Maybe an explosion pushed it in, maybe it squeezed through; it’s hard to say.

But when you have millions of players, one-in-a-million things happen all the time. And those moments are really fun for us, because they show that the system is producing behavior we didn’t explicitly design.

There has been a lot of chatter online about how smart and sophisticated the enemies are, and many players speculate that they're literally learning and evolving by adapting to player habits over time. Is that true?

Martin Singh-Blom: It’s absolutely true that they feel like they’re getting smarter and developing new tactics, but the mechanism is not what people think.

We’re not doing online learning or anything like that. What actually happens is that we see something interesting, maybe in a YouTube video, and then we decide we want more of that behavior.

So we create controlled training scenarios that encourage it. Instead of training on a single fixed setup, we randomize things like object placement so the behavior becomes more robust.

If we only trained on real gameplay data, the behavior could break if something changes in the level. By controlling the training environment, we make it more stable and more generalizable.

Why hasn’t this approach been widely used in games before?

Martin Singh-Blom: Because it is extremely hard.

You’re working at the research frontier of robotics, and it takes years. We almost cut the system multiple times because it didn’t meet our quality bar. There were a lot of discussions about whether we should keep it. The animation didn’t look good enough at certain points, and it was a real risk.

What kept it alive was that the emergent gameplay was so valuable. The game designers really believed in it. Eventually, we implemented a method called adversarial motion priors, and that improved the quality a lot. That’s when we felt confident keeping it.

But it still took a team of five to ten people working for years to get it to this point.

What areas are you most excited to explore further?

Martin Singh-Blom: One interesting area is perception.

Earlier, we had a vision system that couldn’t handle indoor environments. It was basically a grid above the robot that measured the ground height, which worked outdoors but broke indoors.

When we improved that, enemies suddenly became capable of navigating interior spaces, which changed how they behaved and opened up new possibilities. There’s also a lot of potential in expanding the capabilities of existing enemies. Smaller, faster enemies like the leapers are especially interesting because their movement is such a big part of their behavior.

And then there’s a long-term idea of combining this with destruction systems. We’ve experimented with it, and it works, but it would require major changes to the game, so it’s not currently on the roadmap.

Martin Singh-Blom, Machine Learning Research Lead at Embark Studios

Interview conducted by David Jagneaux

Don't forget to subscribe to our Newsletter, join our 80 Level Talent platform, follow us on Twitter, LinkedIn, Telegram, and Instagram, where we share breakdowns, the latest news, awesome artworks, and more.

Are you a fan of what we do here at 80 Level? Then make sure to set us as a Preferred Source on Google to see more of our content in your feed.