Note that when cubes are placed flush against each other, one one of them is detected and the other is ignored.

I think this has something to do with the shadows being part of the training, so the shadowless cube is not detected when two are flush.