Continuous scene perception for accurate world modelling

[alex-mitrevski]

In the current implementation of our perceive_plane action, we store the poses of the detected objects with respect to the base_link frame in the world model, but we don't update them afterwards. This leads to a problem when the robot moves since the objects' poses are not valid anymore - unless we also keep track of the robot's motion. We could alternatively store the poses with respect to a fixed frame, but then we would be relying too much on the localisation being accurate.

An obvious solution here would be reperceiving the scene so that we can correct for any localisation errors when the robot moves. This will take care of some dynamic aspects of the environment (e.g. objects being moved around), but if we want our world model to be accurate, we also need to be able to figure out that we might be looking at the same objects so that we can update their poses accordingly (looking into TUE's wire world model may be worthwhile here).

cc @minhnh