Skip to content
    For Labs

    Every layer of a game world, in trainable form.

    Repos, world packs, tuned environments and ground truth captures. Licensed at the source, delivered the way your lab trains.

    A grid of capture sessions across many licensed worlds
    A slice of the capture library, across 100+ licensed titles.
    Six buyer paths

    Built for how your lab trains

    Select a path, or step through them.

    Action-conditioned pretraining

    Your model learns dynamics from frames paired with the player's true inputs. No latent-action inference, no overlay scraping, no guessing intent.

    inputsclean RGBcamera pose
    Memory that holds up

    Persistent worlds with revisited places. Test object permanence and scene stability over minutes of play, not seconds of clip.

    game stateplayer positionevents
    Honest evaluation

    Held-out worlds that sit outside public corpora. Measure generalization instead of recall, with the same signals you trained on.

    held-out sessions25 configurable modalities
    The action layer: real keystrokes, frame-paired.

    The action layer: real keystrokes, frame-paired.

    100+ licensed titles · 25 configurable modalities · studio-signed provenance

    Start with a sample kit.