Thanks for the great work. This really helps the community!
I wish to point out an issue in the current Pi0FAST implementation. The current implementation constraints the action space to [-1,1] by calling the function normalize_actions https://github.com/huggingface/lerobot/blob/d2645cb19fc521e5b117fe03d90a84f698d3d3f6/src/lerobot/policies/pi0fast/modeling_pi0fast.py#L594.
This normalization is not unnormalized during inference. Moreover, this function is also non-invertible. It normalizes each action chunk by using the minimum and maximum scalar within that action chunk. This information will be different for different action chunks and is not available at inference.