The FAST pipeline operates on action chunks in five steps: (1) Normalization โ raw action chunks are normalized; (2) DCT โ the discrete cosine transform converts normalized actions into frequency-domain coefficients; (3) Quantization โ frequency coefficients are quantized, yielding sparse matrices (most signal energy concentrates in low frequencies); (4) Flattening โ the sparse frequency matrix is flattened with low-frequency components prioritized; (5) BPE โ byte-pair encoding merges frequently co-occurring token pairs into single tokens, typically producing 30โ60 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.
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 โ autoregressive models could not solve them. FAST provides a compact, low-redundancy token representation that enables autoregressive VLAs to learn precise control.
Transforms a normalized action chunk from the time domain to the frequency domain, concentrating signal energy in a few low-frequency coefficients.
Quantizes frequency coefficients into a sparse representation, discarding negligible high-frequency components.
Flattens the sparse coefficient matrix into a sequence, ordering low-frequency components first.
Official
Compresses the flattened sequence by merging frequently co-occurring symbol pairs into single tokens; yields ~30โ60 tokens per chunk.
Official
Autoregressive decoding of 30โ60 tokens per chunk is significantly slower than the flow-matching decode used in ฯ0.
A BPE tokenizer trained on one action distribution may generalize poorly to a different control frequency or action space.
Physical Intelligence releases arXiv:2501.09747, introducing FAST tokenization and ฯ0-FAST โ the first autoregressive generalist policy matching diffusion VLAs at ~5x shorter training.
Release of FAST+ on Hugging Face โ a tokenizer trained on 1M real robot action trajectories that works as a black-box tokenizer across action spaces and control frequencies.
Number of action timesteps encoded in a single chunk before the DCT.
Granularity of DCT coefficient quantization โ controls the trade-off between compression and reconstruction fidelity.
Size of the BPE vocabulary, affecting compression ratio and tokens per chunk.
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 30โ60 tokens, making it slower than ฯ0's flow-matching decode.
The FAST pipeline itself (DCT, quantization, BPE) is lightweight and accelerator-agnostic โ it runs on CPU as preprocessing.
The VLA transformer backbone trained and served on FAST tokens benefits from GPU/TPU tensor cores.
