“The tabletop rearrangement policy was trained on a distribution of 1 to 10 objects sampled from the ShapeNet dataset, which contains approximately 45,000 to 50,000 3D objects.”