Arc Raiders' killer robots aren't learning in real time, but Embark will sneak in AI upgrades based on player behavior: "We keep trying to make the flying drones better and better"

Arc Raiders' use of machine learning in the movements for several of its titular Arc machines has helped fuel conspiracy theories about the robots learning and adapting to player behavior in real time, despite developer Embark Studios repeatedly explaining that this is not how machine learning works in this application. But the folks at Embark, at least, do respond to player behavior and use it to make improvements to the Arc AI – perhaps unannounced.

This has been on my mind ever since I spoke to Arc Raiders machine learning research lead Martin Singh-Blom at GDC, and it's come to the front of my mind following the release of the "Close Scrutiny" map condition that's thrown players into the harshest gauntlet of Arc ever devised. The Arc are never more dangerous than when you're trying to quietly loot silos of rare parts while raider-seeking ordnance launchers whir overhead.

At GDC, Singh-Blom explained how Embark upgraded the Arc's ability to gauge their surroundings in direct response to a popular clip of a Twitch streamer running into a tunnel in order to elude a Leaper, which, at the time, wasn't able to navigate tunnels. It basically couldn't see in such contained spaces, so it avoided them entirely. So, Embark gave it better eyesight under the hood.

Similar upgrades have been made since. Singh-Blom notes that he's only on the machine learning team, not the AI team handling other actions and behaviors, but points to the many flying Arc as one example.

"We've done similar things with the indoor flying behaviors," he says, referring back to the Leaper example. "We keep trying to make the flying drones better and better at flying indoors. And every time players start to think that they're safe in some place, then it's nice if we can figure out a way to make the drones enter that space and make it less safe."

"It's not like we're training on player data," he clarifies. "The loop is, all the game developers look at Tiktoks and YouTube clips and say, hey, that's really cool. Could we make this change that would really make people squeal? And then we make the change, and then people hopefully get surprised and have a good time."

Singh-Blom says, "We hope to have more things in the future and have things in the pipeline, but they're not ready yet," setting up further AI improvements to come. As someone who's cast more than one surprised glance at Hornets and Fireflies that seem to have taken flying courses over the weekend, I can't wait to see how Embark makes me second-guess myself next. Just know that it's a game developer spooking you, not a machine learning on its own.

Arc Raiders has a hidden "magic" system: you've never noticed it because you're not supposed to, but it helps Arc do their job.