FAST - Efficient Robot Action Tokenization

May 26, 2026

  • FAST is an action tokenizer
  • So still in regime of representing actions with discrete tokens
  • Trains 5x faster than diffusion or flow matching
  • You take a normalized action chunk, perform a discrete cosine transform to get the frequency components, then quantize, then convert to a sparse frequency matrix, then flatten by sequencing the lowest frequencies first, then train a BPE tokenizer
  • combining the DCT with BPE yields sequences 10x shorter than otherwise