Action intelligence for game worlds
AI can talk.We make it act.
We build AI that lets game characters understand intent and take physical, strategic, and social action under a world’s physics and rules.
The missing layer is action.
Most AI can generate language or content. Game characters still have narrow, scripted ways to act.
Action intelligence connects intent to behavior. It enables an agent to understand what a person wants, perceive its situation, choose what to do, and carry it out within the constraints of a body, a game, and other agents.
This expands what characters can do beyond a small set of predefined responses and opens the door to game mechanics built around more expressive, adaptive behavior.
From choosing what to say to deciding what to do.
Traditional character behaviorInput → predefined action.
Same input, same authored response. Action intelligenceIntent → world and game state → adaptive action.
Same intent, different resolution as the world changes.
02What we’re building now
Building the ability to move and decide
We are developing two complementary forms of action intelligence: one grounded in bodies and physics, and the other grounded in rules and strategy.
Language → physical action
We are developing physics-based agents that translate language into coordinated full-body movement.
The character must account for balance, contact, momentum, and the changing state of the environment while carrying out an instruction. Rather than selecting only from a fixed animation library, the agent determines how to execute the requested behavior within the constraints of its body and surroundings.
This work can unlock more expressive mechanics involving climbing, gymnastics, athletics, dexterous hand movement, and other behaviors that are difficult to author one movement at a time.
Game state → strategic action
We are building agents for multiplayer board and turn-based games that understand the rules, evaluate possible actions, plan across turns, and adapt to the decisions of other players.
Our longer-term goal is to develop strategic intelligence that can transfer across a family of games rather than requiring a separate hard-coded bot for every title.
From bodies and strategies to teams and worlds
These are not separate product lines. They are stages in a broader effort to build agents that can act across increasingly complex game worlds.
01Bodies
NowFull-body physical control grounded in language, embodiment, and the physics of the environment.
Strategies
NowDecision-making grounded in game rules, world state, possible outcomes, and the behavior of other players.
Teams
NextGroups of agents that interpret higher-level instructions and coordinate their actions.
One example is a soccer team that responds to a coach’s language-based tactical direction rather than requiring the player to control every athlete individually.
Worlds
Longer-term visionRicher simulations populated by NPCs that behave more persistently, socially, and adaptively.
These agents would not merely generate dialogue. They would pursue goals, respond to other characters, make decisions, and produce lasting effects within the world.
Research grounded in playable experiences.
We are both an AI research lab and a game studio.
Our researchers, engineers, designers, and producers work together from the beginning. We do not develop technology in isolation and search for an application afterward. We identify experiences that require a new form of intelligence, build the underlying capability, and turn it into something people can play.
Real games then expose the agents to more varied interactions, unexpected situations, and harder problems—creating the next generation of research questions.
New capability → new game mechanic → real play → harder problems → better agents ↺
Build AI-native games
We create games whose central mechanics are enabled by new forms of character movement, decision-making, and coordination.
Power existing worlds
We work with existing game companies to develop and integrate action intelligence into their characters, gameplay systems, and simulations.
Every game is a new world.
Each game defines a different combination of physics, rules, goals, bodies, action spaces, and social dynamics.
An agent that can adapt across these worlds must learn more than a single task. It must understand what is possible, determine what matters, and choose how to act within a new set of constraints.
Games provide a diverse and controllable distribution of worlds—a major market for action intelligence and a powerful environment for developing increasingly general agents.
Games also allow difficult embodied capabilities to create value before they are practical in robotics.
Climbing, gymnastics, dexterous hand control, musical performance, and other complex behaviors can become compelling game experiences today while advancing methods that may eventually transfer beyond games.