Action normalization in pre/postprocessor
#2
by
xiaoBOSS
- opened
Why is the action normalized with std_mean in the preprocessor, but with identity in the postprocessor? Can someone help to explain this?
The unnormalizer in the post-processor is only applied to action outputs (not visual or state outputs), since the policy (e.g., X-VLA) does not produce image or state logits.
That said, the action post-processor unnormalizer (MEAN_STD / IDENTITY / QUANTILES, etc.) must match the normalizer mode used in the pre-processor of course.
I’ve just updated the base model to fix this.
Note that we generally recommend users choose the normalizer they want at when finetuning the base model.
xiaoBOSS
changed discussion status to
closed