Inconsistent model output across training datasets

For the bridge dataset, the model output is explicitly de-normalized before being compared with the ground truth in the inference notebook.

For other datasets, e.g. Viola, and Berkley cable routing, the raw model predictions match quite well with the ground truth. Does this mean the ground truth actions in these datasets are normalized?

How one should denormalize to interpret the model output e.g. for Franka

It would be helpful to know how datasets are normalized (at least the 9 datasets RT1-X is trained on) and if this is consistent across 9 datasets.

Thanks!