NilsDengler/manipulation_enhanced_map_prediction ? reverse-engineered prompt

I want a Python project that recreates the demo from the paper on manipulation enhanced mapping in a shelf scene. The main thing I need is a runnable simulator where a robot can either take another view or do a push, predict how the map belief would change after candidate pushes, compare the expected information gain, and then choose the better next action. Please make sure the basic shelf environment runs on its own, and that there is also a full end to end demo for the CNABU pipeline using the provided pretrained models if available.

Also include the data collection pieces so I can generate mapping data and push data for later training. If the training and evaluation code from the paper is not fully in this repo yet, that is fine, just make the available parts clean and runnable and note any missing pieces. Please set up installation and any required assets like large model files or submodules, and look up current docs online if you need to.