Adding a camera to an AI system is easy. Giving it a persistent visual world is a different problem.
A naïve system can treat every frame as a new image, repeatedly ask a multimodal model what is visible, and repeatedly search memory against the result. That scales expensive cognition with raw sensor bandwidth — exactly the wrong direction for a continuously exposed system.
The red chair
Imagine LUNA sees a red chair beside a desk. At first the chair is new. After remaining unchanged for hours, it should become computationally quiet without disappearing from the world model.
Someone then stands in front of the chair. A frame-only system might stop detecting it. A persistent perceptual system can instead represent the chair as temporarily occluded. If the camera turns away, the chair becomes out of view rather than nonexistent.
Later the camera returns to the original view. The chair is gone.
09:05 — familiar / stable baseline
09:20 — person blocks view → OCCLUDED
09:30 — camera turns away → OUT OF VIEW
09:40 — camera returns; expected chair absent → MISSING
The disappearance becomes meaningful only because an expectation survived the period in which the chair was not visible. Surprise requires a comparison between the current world and an expected world.
Persist entities, not pixels
LUNA's perception research therefore treats sensory observations as evidence about persistent entities. A frame is evidence about a chair; it is not the chair. An audio segment is evidence that someone spoke; it is not the person.
A compact world model can preserve object identity, location, visibility, familiarity, confidence, spatial relationships, and expectations across time while raw sensor streams remain mostly ephemeral.
Habituation without blindness
Stable input should become less cognitively demanding. The chair can remain strongly represented while contributing almost nothing to current activation because it is familiar, unchanged, and irrelevant to active goals.
A meaningful change resets that quiet state. Movement, disappearance, an unusual relationship, direct attention, or a violated prediction can make the same object salient again.
Different layers should run at different rates
| Layer | Role | Expected cadence |
|---|---|---|
| Sensor acquisition | Capture pixels, audio, depth, motion, or other raw signals. | Hardware rate. |
| Low-level perception | Detect motion, voice activity, tracking primitives, and scene change. | Fast and local. |
| Persistent world model | Maintain entity identity, visibility, familiarity, and expectations. | Event-driven / periodic. |
| Perceptual events | Emit meaningful appeared, moved, missing, spoken, changed, or anomalous events. | Only on meaningful transitions. |
| High-level cognition | Reason, retrieve detail, plan, speak, or act. | Rare relative to raw sensing. |
The compute invariant
High-level cognitive cost should scale primarily with significance and change — not with camera frame rate, microphone sample rate, sensor count, or raw data volume. A physically stable environment should asymptotically approach low cognitive cost even while sensors remain fully active.
Why this matters beyond vision
The same architecture applies to a machine that normally hums and suddenly goes silent, a person expected to arrive but absent, a door usually closed but now open, or an ambient sound that becomes familiar enough to disappear from attention. Prediction error is meaningful because a persistent model of normality exists underneath it.
The research program
A first-class perception layer would make controlled experiments possible on object permanence, habituation, expectation, change blindness analogues, sensory deprivation, cross-modal binding, and the relationship between environmental richness and cognitive activity.
The goal is not to make LUNA constantly narrate what it sees. It is to create an environment that can quietly remain present — until something changes enough to matter.