RESEARCH
Role Playing studies how simulated worlds can be built, entered, and changed — by people and by machines alike. Most systems that work with 3D worlds today reach them indirectly: through a rendered image, or by operating a person's software one action at a time. Our research starts from the opposite premise — that a machine should be able to read a three-dimensional world directly, understand how it behaves, and change it by command. The programs below are the parts of that premise we are actively working on.
Active programs
Authored world state
Can the world a system understands be kept as an explicit, persistent artifact — one that any tool or model can read, operate on, and pass onward — instead of living only as latent state inside the model that computed it?
Today's most capable world models reason in latent space — a compressed internal representation, computed on the fly. It is a superb runtime and the wrong place to keep a world: a latent state is bound to the model that made it, so it does not persist, does not port to another tool, and cannot be read or edited. The representations that do persist — the ones authored and edited by hand — have historically been inert, moved a click at a time. This program joins the two: a model's intelligence operating on an explicit, persistent world directly, by command. A world can be authored by hand or reconstructed from sensors; held that way as the source of truth, through an open, industry-standard description like Universal Scene Description, it becomes durable and directly operable — read and changed by command, not a click at a time.
Runtime-agnostic federation
Can many independent simulation engines — a real-time engine, a high-fidelity physics simulator, an authoring tool — share one world at once, each doing what it does best, without any single engine becoming what everything else depends on?
Every simulator is strong somewhere and weak elsewhere. Federating them around one shared source of world truth — rather than forcing everything through a single runtime — is what lets a world be interactive and physically accurate at once, and keeps the work from being captive to one engine or vendor.
Portable replays
Can a recording of what happened in a world carry its meaning with it — replayable and interrogable in a different system than the one that produced it?
The 3D world already has portable recordings: a baked geometry cache like Alembic moves between tools well enough. But it arrives flat — you get where everything was, frame by frame, and nothing else. A replay that also carries its causal structure could be reviewed after the fact, used to train a model, or handed to another team. That portability of meaning, not just geometry, is what turns a recording into an asset you can reuse and build on.
Direction
These questions converge on one capability: a system capable of authoring, manipulating, and understanding a world from an embodied entity within it or from an omniscient view outside it. The same substrate serves a crisis-response rehearsed before it occurs, a machine tested before it ships, and a story role-played rather than filmed. The long-term work is to make that system real, and good enough that the difference between those uses is a matter of intent rather than of technology.