The shelves are composed of several parts, and we have meshes in different levels of complexity.

Finding a single reference frame for the parts of the shelves is hard because different actors that deal with the meshes have different needs:

• modelling: Prefer to have the reference frame at the center of the bounding box (can be automated)
• perception: prefers a reference frame placed on an easily detectable part of the object
• building the shelf from perceived data: would prefer a reference frame at the back of the shelf parts

So after discussing this with Andrei, Georg, and Alexis, our current proposal is the following:

• Shelf-parts get modelled and the frame of reference (origin) gets placed at the center of the bounding box.
• The orientation of that reference frame is chosen such that it is identity when placed on the real world. In practical terms, all the shelf floors will have the same orientation as the shelf system.
• We will extend the object-state-publisher to publish static-transforms for additionally defined frames that are useful for the other purposes

We will use a corner of the room, because it is an unmovable feature that we can find in the laser-scans, and we can measure from it in real life.