Embark trains neural networks using reinforcement learning to control physics-based character locomotion, replacing traditional hand-animated state machines. Robot enemies learn to walk, balance, and react to damage emergently — if you shoot off a leg, the NPC attempts to rebalance on its own.

This is Embark's most technically ambitious and proudly showcased AI capability. A dedicated team of ~12 engineers, researchers, tech artists, designers, and animators has worked on this since at least 2019. The system uses Soft Actor-Critic (SAC) reinforcement learning algorithms — agents observe their body and surroundings, then decide frame-by-frame how to move. Tom Solberg, ML Software Engineer, wrote in 2021: "The role of our animators is no longer to draw animation curves… they're here to train agents." The system produces emergent behaviors impossible with traditional animation — CCO Strandberg confirmed "if you shoot off a leg, it'll try to rebalance." Embark open-sourced two key components: emote (their RL training library, Python/PyTorch) and cervo (Rust middleware for ML inference in games). At GDC 2022, Jorge del Val Santos presented "Walk Lizzie, Walk!" at the Machine Learning Summit, and at GDC 2026, Martin Singh-Blom presented on the production implementation in ARC Raiders alongside panelists from EA SEED, Ubisoft La Forge, and Google DeepMind.