Memory that stores grid models, which is what an evidence-based learning module learns into.

Mirrors tbp/monty/frameworks/models/evidence_matching/graph_memory.py ::EvidenceGraphMemory, whose whole content is a different object model plus what follows from it. Evidence matching compares one observation against every node of every object at once, so memory that stored every observation as a node would be both slow and enormous; the grid keeps a bounded number of nodes per object and is the reason a pretrained YCB model fits in memory at all.

Inheritance became configuration

In Python this class exists to subclass GraphMemory and override the three places where a model is constructed — build, extend, and load. There is no inheritance to override in Elixir, and there does not need to be: memory already takes the object model as a module and its options as a keyword list, so the same three places build a GridObjectModel when memory is made here and a GraphObjectModel when it is made by GraphMemory.new/1 directly. One struct, one set of functions, two configurations.

What the three numbers mean

• :max_nodes_per_graph — how many nodes an object thins down to, which bounds the cost of every matching step;
• :max_graph_size — how large an object may be, in metres, before the model refuses it with GridTooSmallError;
• :num_model_voxels_per_dim — the resolution, hence the smallest detail a model can hold: max_graph_size / num_model_voxels_per_dim per voxel.