Does Arc Raiders use machine learning for its enemies?

Yes, but only for movement. Embark used reinforcement learning borrowed from robotics research to teach the legged ARC how to place their feet, balance, climb uneven terrain, and recover after being knocked over. Everything else, including detection, aggro, attacks, and the decision of where to go, is hand-authored game AI running on behavior trees and utility AI. Flying drones do not use it at all, since Embark uses traditional control systems for them.

Does Embark train the ARC on player data?

No. Singh-Blom told GamesRadar at GDC, "It's not like we're training on player data." New behaviors come from developers watching clips players post online, then building controlled training scenarios offline with randomized conditions. He told 80.lv they deliberately avoid training on real gameplay, because behavior learned from live play breaks when the level changes.

Why do the ARC feel like they are getting smarter?

Because their movement is physically simulated rather than animation locked, so they stumble, get knocked off course, recover, and end up in places nobody scripted. Physics also produces slightly different results every run, so no two fights repeat. On top of that, Embark really does improve the ARC over time, most notably with a perception upgrade that let them navigate tunnels and interiors they used to refuse to enter. All of it reads as intent, but none of it is a machine adapting to you personally.

Which ARC enemies use the machine learning locomotion?

The rule Embark has stated publicly is legged versus flying: multi-legged ARC use machine-learned locomotion, and flying drones use traditional control systems like real-world drones. Leapers and Bastions are the machines named across multiple reports of Embark's GDC talk. Embark has not published a machine by machine list, so any complete roster you see elsewhere is inference rather than confirmation.

Is the ARC enemy AI a large language model or generative AI?

No. Asked by PC Gamer whether any AI-related tech drives ARC behavior, design director Virgil Watkins answered "None whatsoever," clarifying that he meant AI in the game development sense, meaning hand-authored enemy design. The locomotion uses reinforcement learning, a robotics technique, not a generative model. Arc Raiders did face separate criticism over text-to-speech NPC voices, later replaced with human-recorded dialogue in many lines, but that is audio production and unrelated to enemy AI.

Did the ARC get smarter after launch?

Their decision making did not change, but where they can physically go did. Game Rant reported in February 2026 that ARC enemies are the same as they were at launch, with post-launch updates aimed at pathfinding and bugs, "but some of them can just walk a little better now." The biggest real change was perception: Embark replaced an overhead height-map vision system with point clouds cast from the robot's own viewpoint, which let ARC follow players into tunnels and interiors that used to be safe.

Arc Raiders Machine Learning: How the ARC Enemy AI Really Works

The ARC do not learn your habits mid raid. Embark used reinforcement learning to teach them to walk, balance and climb. Here is what that changes in a fight.

Category: Arc Raiders AI & ARC Machines · Updated 2026-08-03

Search "arc raiders machine learning" and a popular theory comes up: the ARC watch how you play, learn your habits, and adapt to counter you. That is not happening. Embark Studios did use machine learning on these machines, but it taught them how to move, not how to read you.

The short answer

PC Gamer asked design director Virgil Watkins directly whether the ARC are learning new techniques. His answer: "None whatsoever... The machine learning is literally only for teaching them to walk and navigate the environment. It doesn't do any of their behaviours or their attacks or anything like that."

Martin Singh-Blom, Embark's machine learning research lead, said the same to 80.lv: "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." At GDC he told GamesRadar: "It's not like we're training on player data."

Nothing you do in a raid updates anything. The model that ships is frozen. PC Gamer traces the myth to a 2021 post from Embark by machine learning software engineer Tom Solberg about teaching enemies to move and navigate organically instead of hand animating them. Game Rant points at the same 2021 Embark post without naming its author. It was about locomotion. It got read as a claim about intelligence.

What the machine learning actually does

Singh-Blom, to 80.lv: "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."

The rule for which machines use it is legged versus flying. Drones do not: "For drones, we don't need that since we can use traditional control systems, as you'd see in real-world drones." Leapers and Bastions are the examples named in reports of the GDC talk. Other legged ARC plausibly qualify, but Embark has never gone machine by machine in public, so treat any full roster elsewhere as inference.

Years, not minutes

The loop: a policy network outputs actions, the simulated robot performs them in physics, observes the result, and a reward nudges the policy toward whatever scores higher. Singh-Blom compared it at GDC to teaching a dog a trick. Early setups rewarded closing on a target and staying upright.

Walking without falling was not the hard part. Walking convincingly was. The first working gait had no weight to it and no living rhythm, and the core problem is that speed and turn rate are measurable while "does the walk look cool" is not. As reported from the GDC talk, the animation director watched endless generated video and handed back notes like "it's a bit jerky," leaving the engineers nothing to optimize against.

How long did that take? "It takes years," Singh-Blom said, and "it still took a team of five to ten people working for years to get it to this point." No published source gives an iteration count, so ignore the "millions of attempts" line in aggregator posts. The system nearly got cut for missing Embark's quality bar; what saved it was that "the emergent gameplay was so valuable."

The fix was adversarial motion priors. An animator hand-keys a reference walk in the ordinary curve editor, and a second network has to tell that animation apart from the robot's own motion. The robot earns extra reward when it fools it, so changing how a machine walks means making a new reference animation, no ML knowledge required. Training one network on several tasks made them interfere, so each machine ended up with several brains, one policy network per task, with an ordinary behavior tree switching between them. Reporting the GDC talk, GamesRadar describes a "patrol walk brain" and a "pursuit brain."

Perception, which is what actually changed

This is the most useful fact for anyone who thinks the ARC learned their hiding spot. The old vision system was a grid of points above the robot casting rays down to build a height map. It worked outdoors and broke indoors: rays hit tunnel ceilings and read them as obstacles, so machines refused to enter. That broke in public at the Server Slam test, when a downed player crawled into a tunnel and the pursuing robot bounced out of a gap it should have fit through.

The fix was point clouds: rays cast from the robot's own viewpoint in all directions, processed with PointNet. Singh-Blom: "When we improved that, enemies suddenly became capable of navigating interior spaces, which changed how they behaved and opened up new possibilities."

That is the "it learned where I hide" moment: a shipped perception upgrade. It changed where the ARC can go, not what they decide to do.

Why it still feels like learning

Movement is not animation-locked, so the ARC end up where nobody scripted. Singh-Blom describes an internal video of a Leaper inside a dome: "we didn't even know it could get in there." With 16 million copies sold and over 1.5 billion hours played, one-in-a-million outcomes happen constantly, and each looks like intent.
Physics gives slightly different results every run, so no two fights repeat.
Embark does improve the ARC, but backwards to the myth. Developers watch player clips, then build controlled training scenarios offline, randomizing conditions so the behavior holds up. They avoid training on live play, because behavior learned that way breaks when the level changes.
The ordinary AI is simply good. Singh-Blom: "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."

What this changes at the moment of contact

They stumble for real. A Leaper can lunge, botch the landing, tumble, and get back up. Not a canned animation.
They can be moved. "If you throw something at it, it might get bumped out of its path." Shoot a machine into a wall and, per Singh-Blom, it can take extra damage and explode from the impact rather than from your gun.
De-limbing degrades, it does not disable. A capped assist force called magic torque keeps a Leaper or Bastion moving after losing multiple legs, so do not count a severed leg as a mobility kill.
Damaged machines still want to stand. Creative director Stefan Strandberg, in the studio docuseries: "when a machine loses its limbs, it still wants to stand up." They recover from impact rather than resetting.
Terrain is not a reliable wall, and since the perception upgrade neither are tunnels and interiors. No geometry is safe on the theory that a designer must have thought of it.
Read intent instead of memorizing patterns. That is the designers' goal: "It shifts the game away from memorizing patterns and toward understanding behavior."
None of it is personal. The same machine would do the same spread of things to anyone in the same situation, and nothing carries between raids.

Tactical claims in guide coverage, such as ARC flanking or repositioning after losing a leg, are journalists' play observations. They fit the systems above but are not dev-confirmed.

Machine learning versus ordinary game AI

IS machine learning:

Locomotion, "how the robot places its feet and moves"
Legged ARC only. Leaper and Bastion are the confirmed examples
The gait, via adversarial motion priors against an animator's reference
Balance, staying upright, recovering after impact
Terrain traversal: climbing, uneven ground, going over an obstacle or around
Local terrain perception (point cloud, PointNet) feeding the locomotion policy
Whether to spend magic torque when crippled, which sits in the reward scheme

Is NOT machine learning:

Every behavior and every attack. Watkins: "It doesn't do any of their behaviours or their attacks or anything like that."
Higher-level decisions about where to go and what to do, and which brain is driving. Behavior trees.
Detection and aggro. An ordinary state machine, and the scanner light tells you the state: blue for passive, yellow for alerted, red for active engagement.
Flying drone movement, which uses conventional control systems.
Enemy tuning, done with utility AI by Embark's AI team.
Any adaptation to you at runtime. There is none.
Any LLM or generative AI in enemy behavior.

Singh-Blom's boundary line is the cleanest summary of it: "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." He has said that boundary may move as models improve, but that is the future, not the game you play today.

The separate criticism over text-to-speech NPC voices, later replaced with human-recorded dialogue in many lines, is an audio production matter, not enemy AI. The sourcing above is Embark developers speaking publicly plus first-hand reports of their talks: PC Gamer's interview with Watkins, 80.lv's with Singh-Blom, 4Gamer's reports from his GDC 2026 and NDC26 talks, and GamesRadar's GDC coverage. Where a phrase is a reporter's wording rather than a developer's, such as the brain names, it is attributed that way above. This is a fan-made companion site, not affiliated with Embark Studios.

Frequently asked questions

Do the ARC in Arc Raiders learn your playstyle?: No. The ARC do not adapt to individual players, and nothing you do in a raid changes how they behave later. Embark design director Virgil Watkins told PC Gamer the machine learning is "literally only for teaching them to walk and navigate the environment" and that it does not touch their behaviours or their attacks. Machine learning research lead Martin Singh-Blom put it plainly to 80.lv: "We're not doing online learning or anything like that."
Does Arc Raiders use machine learning for its enemies?: Yes, but only for movement. Embark used reinforcement learning borrowed from robotics research to teach the legged ARC how to place their feet, balance, climb uneven terrain, and recover after being knocked over. Everything else, including detection, aggro, attacks, and the decision of where to go, is hand-authored game AI running on behavior trees and utility AI. Flying drones do not use it at all, since Embark uses traditional control systems for them.
Does Embark train the ARC on player data?: No. Singh-Blom told GamesRadar at GDC, "It's not like we're training on player data." New behaviors come from developers watching clips players post online, then building controlled training scenarios offline with randomized conditions. He told 80.lv they deliberately avoid training on real gameplay, because behavior learned from live play breaks when the level changes.
Why do the ARC feel like they are getting smarter?: Because their movement is physically simulated rather than animation locked, so they stumble, get knocked off course, recover, and end up in places nobody scripted. Physics also produces slightly different results every run, so no two fights repeat. On top of that, Embark really does improve the ARC over time, most notably with a perception upgrade that let them navigate tunnels and interiors they used to refuse to enter. All of it reads as intent, but none of it is a machine adapting to you personally.
Which ARC enemies use the machine learning locomotion?: The rule Embark has stated publicly is legged versus flying: multi-legged ARC use machine-learned locomotion, and flying drones use traditional control systems like real-world drones. Leapers and Bastions are the machines named across multiple reports of Embark's GDC talk. Embark has not published a machine by machine list, so any complete roster you see elsewhere is inference rather than confirmation.
Is the ARC enemy AI a large language model or generative AI?: No. Asked by PC Gamer whether any AI-related tech drives ARC behavior, design director Virgil Watkins answered "None whatsoever," clarifying that he meant AI in the game development sense, meaning hand-authored enemy design. The locomotion uses reinforcement learning, a robotics technique, not a generative model. Arc Raiders did face separate criticism over text-to-speech NPC voices, later replaced with human-recorded dialogue in many lines, but that is audio production and unrelated to enemy AI.
Did the ARC get smarter after launch?: Their decision making did not change, but where they can physically go did. Game Rant reported in February 2026 that ARC enemies are the same as they were at launch, with post-launch updates aimed at pathfinding and bugs, "but some of them can just walk a little better now." The biggest real change was perception: Embark replaced an overhead height-map vision system with point clouds cast from the robot's own viewpoint, which let ARC follow players into tunnels and interiors that used to be safe.