FAST: Frequency-Space Action Sequence Tokenization for VLA Action Generation
FAST (Frequency-space Action Sequence Tokenization) tokenizes continuous robot action sequences by converting each action dimension from time to frequency using the Discrete Cosine Transform (DCT), sparsifying/rounding insignificant coefficients, then flattening them into a discrete token sequence. Byte Pair Encoding (BPE) is applied to reduce redundancy further, and the process is described as invertible so actions can be reconstructed from tokens.
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
In Vision-Language-Action (VLA) models, how actions are represented affects efficiency, fidelity of executed motions, and generalization across different robot embodiments. FAST is presented as an alternative to common discrete per-dimension/per-timestep binning tokenizers, aiming to produce a more information-dense and less redundant action token sequence.
What changed
Instead of discretizing each continuous control value directly into bins over time, FAST transforms action trajectories into the frequency domain with DCT, removes/sparsifies small coefficients via scaling-and-rounding, then tokenizes the flattened coefficient sequence with BPE.
Where it helps
- Action-tokenizers for vision-language-action robot policies
- Autoregressive VLA models that need compact action representations
- Cross-embodiment robot control where action sequences vary in morphology
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
The approach introduces hyperparameters controlling the DCT coefficient scaling before rounding and the BPE vocabulary size; the source text does not provide how sensitive training outcomes are to these choices beyond stating that parameters are described as robust.
Sources
- π0 and π0-FAST: Vision-Language-Action Models for General Robot Control ↗
Hugging Face · primary · Published 2025-02-04 · Checked 2026-11-10
