Generalization to box sizes outside of those seen during training. L conveys the X- and Y- dimensions of the environment (Z-dimension has length 3 for all boxes). Simulators shown include the ground truth (first row) as well as the learned Dilated ResNet models trained on data with L=0.75 (second rows) and L in {0.5, 0.75, 1, 1.25} (third row). Generalization values of L range from L=0.25 to L=2 in increments of 0.25.

We found that models trained on multiple sizes generalize better to sizes outside the training distribution.