The World-Model Stack

Perception Backbones · pixels → features & detections

YOLO DINOv2 V-JEPA 2 SAM VLMs

any callable(image) → [Detection]

detections + embeddings

The Encoder & Standardized State

Trio Retina

one model-agnostic WorldState — symbolic events + latent state s = Enc(o)

the invariant interface between perception and dynamics

one standardized state stream

World Models · dynamics & control · state → prediction → action

latent dynamics models p(sₜ₊₁ | sₜ, a) forecasting

model-based RL & planning · latent rollouts · policy · agents / digital twins

Swap any encoder above or any dynamics model below — Retina is the invariant state in the middle.