A camera result is evidence, not decoration#

Ainekio does not continuously stream room video into a model. The physical camera path is designed around fresh, bounded still images tied to a reason: an explicit snapshot request, a completed motion action, or a VAD-bounded utterance.

The operator dashboard can request a low-rate preview for human inspection, but preview remains off until explicitly enabled. Model-facing perception uses the correlated still path.

Camera behavior on the robot#

The OV3660 controller captures fresh XGA JPEG stills. Preview is limited to QVGA or VGA at a bounded requested rate. One PSRAM framebuffer is used, and the capture engine is idle between frames.

This keeps the camera from becoming an always-on bandwidth and memory owner. It also makes the source of an observation explicit.

The current fast capture profile favors the first useful frame after motion:

short manual exposure aligned with common indoor 60 Hz lighting;
moderate gain rather than maximum brightness at any cost;
denoise with neutral sharpening;
no slow automatic exposure convergence;
JPEG quality chosen to preserve detail inside the bounded payload.

That profile came from physical evidence. An earlier high-gain and sharpened profile produced severe cross-hatch artifacts, so the repair reduced gain and removed sharpening instead of assuming a brighter signal was always better.

Three snapshot origins#

Explicit request#

MetaHuman or the dashboard asks for a fresh frame. The robot publishes camera metadata with a request origin and sends the bounded JPEG.

Action completion#

After a completed motion, the controller can capture a frame with the command sequence as its origin identifier. The JPEG is sent before the final done so the Environment path can associate the physical result with the action that preceded it.

Audio utterance#

When VAD closes an utterance, the controller can capture a frame with the same robot-generated utterance identifier. The gateway and Environment Bridge can join what was heard with what was visible at that event boundary.

The return path#

OV3660 fresh JPEG
  -> protocol camera metadata and binary frame
  -> Ainekio gateway
  -> Environment adapter observation
  -> MetaHuman Environment Bridge
  -> current correlated Environment context
  -> Environment Task State

The Environment context carries frame identity, action or utterance correlation, capabilities, and current body state. A visual selector can therefore reason about a specific frame rather than whatever image happens to be latest in a mutable cache.

Action completion versus visual completion#

Consider: “Walk until there is enough light.”

If the task closes on action_result, a successful walk only proves that the robot walked. It does not prove that the new location is brighter. The stopping condition depends on a fresh view, so the task must require visual_observation.

Environment Task State now keeps those bases distinct. When a fresh frame does not prove the stopping condition, bounded continuation should preserve the previously admitted action rather than silently substituting a different action or declaring success.

This is one of the project’s most important semantic boundaries:

Terminal robot success answers “did the command finish?” Visual evidence answers “did the world now satisfy the objective?”

What the visual system is not yet#

Ainekio does not currently advertise a general target-relative navigation or visual-approach capability. A named findObject concept elsewhere in MetaHuman OS does not make the body capable of locating and approaching an object.

Earlier Active View research explored detection and temporal segmentation, but it remains isolated from the maintained production path. Detection continuity cannot create semantic identity or completion authority on its own.

Next visual acceptance#

The next meaningful demonstration is a repeated physical objective with all of these captured together:

admitted Environment task and required completion basis;
dispatched action identifier;
robot acknowledgement and terminal result;
fresh frame with matching correlation;
frame-cited decision about the stopping condition;
bounded continuation when the condition is absent;
final completion only when a later fresh frame proves it.

That is more demanding than showing a camera preview, and much closer to what an embodied companion actually needs.