FAST: Efficient Action Tokenization for Vision-Language-Action Models

This is quite important.

How do robot tokenizers work?