Turns one camera observation into a CMP message: a location, a surface, and the features a learning module matches on.

Mirrors tbp/monty/frameworks/models/sensor_modules.py::ObservationProcessor. It is the pure part of CameraSM — an observation in, a percept out, no state — which is why Monty has it as a class of its own too.

What it reads

The observation has been through the experiment's transforms, so it carries "semantic_3d" (one row per pixel: three coordinates and the on-surface mask), "cam_to_world", "rgba" and "depth". The middle pixel is half + width * half of the point cloud — Monty's own arithmetic for a square patch — and on_object is that row's mask.

Which features

:features is the list a config declares, and a feature not in it is not computed. That is the point of declaring it: feature_weights in a hidden layer and the tolerances in graph matching are written against exactly these names, and a percept carrying more would be one Monty's modules never see. :object_coverage, :min_depth and :mean_depth are taken over the pixels on the surface rather than over the whole patch.

One difference from Monty

Monty reads features["object_coverage"] when a surface sensor's middle pixel is off the object, and raises KeyError if the config did not ask for that feature. Here a missing coverage counts as zero (docs/monty-differences.md §2.10).