Robotics

FAST

2025ActivePublished: 1 September 2026Updated: 1 September 2026Published

Key innovation

Instead of naive per-dimension, per-timestep binning, FAST compresses continuous action sequences in the frequency domain (DCT + quantization + BPE), letting autoregressive VLA models learn dexterous, high-frequency tasks that standard discretization schemes failed to solve.

How it works

The FAST pipeline operates on action chunks in five steps: (1) Normalization — raw action chunks are normalized; (2) DCT — the discrete cosine transform converts normalized actions into frequency-domain coefficients; (3) Quantization — frequency coefficients are quantized, yielding sparse matrices (most signal energy concentrates in low frequencies); (4) Flattening — the sparse frequency matrix is flattened with low-frequency components prioritized; (5) BPE — byte-pair encoding merges frequently co-occurring token pairs into single tokens, typically producing 30–60 tokens per chunk. The resulting tokens are predicted autoregressively by a transformer (the VLA backbone), and at decode time the pipeline is inverted to reconstruct continuous actions.

Problem solved

Standard action tokenization (discrete binning of each dimension at each timestep) produces long, highly correlated token sequences and fails completely on dexterous, high-frequency tasks — autoregressive models could not solve them. FAST provides a compact, low-redundancy token representation that enables autoregressive VLAs to learn precise control.

Components

Discrete Cosine Transform (DCT)Decorrelation and compression of the action signal

Transforms a normalized action chunk from the time domain to the frequency domain, concentrating signal energy in a few low-frequency coefficients.

QuantizationPrecision reduction and sparsity

Quantizes frequency coefficients into a sparse representation, discarding negligible high-frequency components.

Flattening (low-frequency first)Serialization into a token sequence

Flattens the sparse coefficient matrix into a sequence, ordering low-frequency components first.

Official

Byte-Pair Encoding (BPE)Vocabulary-based compression of the token sequence

Compresses the flattened sequence by merging frequently co-occurring symbol pairs into single tokens; yields ~30–60 tokens per chunk.

Official

Implementation

Reference implementations

FAST+ tokenizer (Hugging Face)

Python · Physical Intelligence

Official

openpi (π0 / π0-FAST)

Python (JAX) · Physical Intelligence

Official

Implementation pitfalls

Slower inference than flow matchingMedium

Autoregressive decoding of 30–60 tokens per chunk is significantly slower than the flow-matching decode used in π0.

Fix:Use when fast training / autoregressive simplicity is the priority; for low-latency inference consider flow-matching variants.

Tokenizer fit to the action spaceMedium

A BPE tokenizer trained on one action distribution may generalize poorly to a different control frequency or action space.

Fix:Use the universal FAST+ (1M trajectories) or retrain the tokenizer on the target action distribution.

Evolution

Original paper · 2025 · arXiv (2025) · Karl Pertsch

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

Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, Sergey Levine

2025

FAST and π0-FAST released

Inflection point

Physical Intelligence releases arXiv:2501.09747, introducing FAST tokenization and π0-FAST — the first autoregressive generalist policy matching diffusion VLAs at ~5x shorter training.

FAST: Efficient Action Tokenization for Vision-Language-Action Models (paper)

2025

FAST+ — universal action tokenizer

Release of FAST+ on Hugging Face — a tokenizer trained on 1M real robot action trajectories that works as a black-box tokenizer across action spaces and control frequencies.

Hyperparameters (configurable axes)

Action chunk sizeHigh

Number of action timesteps encoded in a single chunk before the DCT.

Quantization scaleHigh

Granularity of DCT coefficient quantization — controls the trade-off between compression and reconstruction fidelity.

BPE vocabulary sizeHigh

Size of the BPE vocabulary, affecting compression ratio and tokens per chunk.

Parallelism

Parallelism level

Partially parallel

FAST encoding (DCT+quantization+BPE) is cheap and done offline/on CPU. Autoregressive training over tokens is parallel (teacher forcing), whereas inference requires sequential autoregressive decoding of 30–60 tokens, making it slower than π0's flow-matching decode.

Scope

Training