Embark Studios uses physics-driven ML to power unpredictable enemy AI in Arc Raiders
Summary
Embark Studios built Arc Raiders' enemy AI on physics simulation and reinforcement learning, borrowed from robotics research
Enemies are treated as physical entities rather than scripted characters, creating emergent and unpredictable behavior
Machine learning is used strictly for locomotion in legged robots, while behavior trees handle high-level decisions
The system took a team of 5-10 people several years to develop and was nearly cut multiple times
"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"
01
Physics as the foundation - not just a visual layer
All enemy behavior flows from a core decision to treat enemies as physical objects, not animated characters
Traditional animation locks enemies to fixed positions, while physics means enemies can be displaced, pushed, and disrupted
This unpredictability forced Embark to abandon rigid pathfinding in favor of flexible, self-correcting systems
"Since we decided to go with physics, we don't necessarily know where the enemy will be"
Emergent interactions are by design - shooting an enemy into a wall deals extra damage from impact, not just the weapon
02
How ML fits in - narrower than players assume
Reinforcement learning is used exclusively for locomotion: how legged robots place feet and move through space
Drones use traditional control systems; only legged enemies require ML-based movement
High-level decisions such as where to go and what to do are handled by behavior trees, not machine learning
A key boundary exists: the behavior tree decides intent, the locomotion system decides execution
As ML models improve, more decision-making can shift to the learned side, increasing unpredictability
03
Emergent behavior - the system surprises even the developers
Enemies have appeared in locations developers never anticipated, with no clear explanation of how they got there
"Those moments are really fun for us, because they show that the system is producing behavior we didn't explicitly design"
At scale with millions of players, one-in-a-million outcomes happen constantly, keeping gameplay fresh
Enemies are not doing online learning or adapting in real time to players, despite widespread player belief
Observed interesting behaviors are manually recreated in controlled training scenarios to make them more robust and repeatable
04
Why this approach is rare - years of risk and near-cancellation
The system sits at the research frontier of robotics and required years of sustained investment
The project was nearly cut multiple times when animation quality did not meet the studio's standards
A breakthrough came with the adoption of adversarial motion priors, which significantly improved visual fidelity
"It still took a team of five to ten people working for years to get it to this point"
Game designers' belief in the emergent gameplay value was the key reason the system survived
05
What Embark is exploring next
Perception systems are a key focus - improved indoor navigation opened up entirely new enemy behavior possibilities
Smaller, faster enemies like leapers are a priority given how central movement is to their design identity
Long-term research includes combining physics-driven AI with destruction systems, though this is not yet on the roadmap
The broader direction is pushing more decision-making into the ML layer to create increasingly surprising player encounters
Events
Companies
Games
Locations
—
COMPILED BY GAMES ATLAS EDITORIAL
RELATED ARTICLES
EXPLORE MORE
Comments (0)
No comments yet. Be the first to share your thoughts!