Where FAST came from, and what it led to

FAST replaced the per-dimension binning of RT-2 and OpenVLA with JPEG-style compression of action chunks. Physical Intelligence’s later models, π0.5 and π*0.6, still train on FAST tokens but decode actions with flow matching, while other groups argue for fixed-length tokens, learned tokenizers or no tokenizer at all.

Paper: FAST, arXiv:2501.09747 Data: Semantic Scholar, 19 Oct 2026 Role: archeologist October 2026

Citation timeline

72 references · 141 of 781 citing papers

Every paper FAST cites sits to the left of January 2025; papers that cite FAST sit to the right. Higher means more cited. A ring marks the papers discussed below. Pick chips under the search box to show only those papers; All shows everything, and FAST always stays. Click a paper to draw all its links to the others; for a citing paper the panel quotes the sentences it uses to cite FAST.

Loading the citation graph…

Left to right: publication date, stretched near 2025. Up: citations, log scale. Dot size: citations. Details for the selected paper appear below the graph.

Prior works

Before FAST

  1. 1974–2016
    The DCT, JPEG and byte-pair encoding
    Shaped FAST · refs. [1], [61], [27], [57]
    The idea comes from outside robotics. Take the discrete cosine transform of each action dimension, scale and round so small high-frequency terms vanish, as JPEG does for image blocks, then merge frequent patterns with BPE, as language models do for text. A one-second shirt-folding chunk at 50 Hz goes from 700 binning tokens to 53 FAST tokens.
  2. 2017–2024
    Learned quantizers: VQ-VAE, FSQ, VQ-BeT, MiniVLA
    Shaped FAST · refs. [59], [48], [41], [3]
    The road FAST did not take. FAST finds that these work for coarse reconstruction but fail on high-frequency tasks that need fine control; FSQ is its learned-compression ablation.
  3. Apr 2023
    ACT and ALOHA
    Shaped FAST · ref. [69]
    Action chunking: predict a second of future actions at once. This makes long, high-frequency chunks the target that a tokenizer has to handle.
  4. Jul 2023
    RT-2
    Shaped FAST · Google DeepMind · ref. [10], cited 11 times
    The first VLA. It turns each action dimension at each time step into one of 256 bin tokens. FAST’s case study shows why this breaks at high control rates: consecutive tokens are almost identical, so the next-token loss is low when the model just copies the last token.
  5. Mar 2024
    DROID
    Shaped FAST · ref. [38], cited 11 times
    A large in-the-wild manipulation dataset that binning-based autoregressive VLAs could not train on well. FAST calls the π0-FAST policy trained on it the first language-conditioned generalist manipulation policy that works zero-shot in unseen environments.
  6. Jun 2024
    OpenVLA
    Shaped FAST · ref. [39], cited 18 times
    An open 7B VLA that kept RT-2’s binning, and FAST’s main baseline (“OpenVLA-style” in its Figure 2). FAST’s first author, Karl Pertsch, is a co-author of RT-2, OpenVLA, Octo and DROID, so FAST fixes a limit of his own earlier models.
  7. Oct 2024
    π0: a vision-language-action flow model
    Shaped FAST · Physical Intelligence, the same lab · ref. [7], cited 28 times, more than any other
    The lab’s earlier model and FAST’s base. π0 is PaliGemma-3B with a flow-matching action expert; π0-FAST keeps the backbone, drops the expert and writes FAST tokens into the VLM’s vocabulary, and the paper compares the two decoders. π0-FAST matches π0’s success with 5× fewer GPU hours, but needs about 750 ms per one-second chunk against about 100 ms, because it decodes 30 to 60 tokens with the 2B backbone instead of 10 denoising steps in a 300M expert.

What came after

Later work that cites FAST

  1. Feb 2025
    OpenVLA-OFT
    Challenges · Kim, Finn and Liang · 1,088 citations
    “Significant latency between action chunks (e.g., 750 ms for the recent FAST approach) still limits real-time deployment on high-frequency bimanual robots.” It drops tokens altogether, with parallel decoding, continuous actions and an L1 loss, and reports higher success than the best fine-tuned π0 policies, π0-FAST included.
  2. Apr 2025
    π0.5
    Validates · Physical Intelligence · 2,415 citations, the most of any paper citing FAST
    All pre-training data, robot actions included, is discrete tokens via FAST, “which leads to simple, scalable, and efficient training”; flow matching produces the actions at inference. Eight of FAST’s nine authors are on it.
  3. May 2025
    Knowledge Insulating VLA models
    Extends · Driess, Pertsch, Levine and others, Physical Intelligence
    Trains the VLM on FAST tokens and a flow-matching action expert at the same time, with a stop-gradient so the expert cannot damage the VLM: FAST’s training speed with π0’s inference speed. FAST’s own discussion lists this as future work, “the combination of compression-based action encodings with non-autoregressive decoding approaches like diffusion”.
  4. Jun 2025
    BEAST
    Challenges · B-spline action tokenizer
    FAST gives “variable-length token sequences for inputs of the same duration, which complicates applying fast token generation techniques such as parallel decoding.” B-spline control points give fixed-length, smooth tokens and need no tokenizer training.
  5. Jun 2025
    Real-time chunking
    Extends · Physical Intelligence
    Removes the pause between action chunks. It cites FAST among the prior work that stops the robot while the next chunk is generated.
  6. Oct 2025
    VLA-0
    Challenges · actions as plain text
    Writes actions as plain text numbers with no custom tokenizer, calls π0-FAST a “custom” VLA, and reports higher LIBERO success than OpenVLA-OFT, π0.5 with Knowledge Insulation and SmolVLA.
  7. Nov 2025
    π*0.6
    Validates · Physical Intelligence
    Still predicts FAST tokens during training, through the Knowledge Insulation recipe.
  8. Dec 2025
    FASTer
    Extends · learned action tokenizer
    Replaces the DCT with a learned VQ tokenizer that treats an action chunk as an image, and adds block-wise autoregressive decoding with a small action expert.
  9. Feb 2026
    ActionCodec and OAT
    Challenges · tokenizer design principles
    ActionCodec says FAST’s “reliance on fixed geometric priors limits its capacity for cross-embodiment knowledge transfer” and that it “requires a specialized sequence-length alignment heuristic”. OAT asks for tokenizers that compress, always decode and are ordered left to right, and learns one that does all three.
  10. Apr 2026
    The Compression Gap
    Challenges · Shiba
    Argues that a discrete action codebook caps how much a better vision encoder can help, and notes that FAST “suffers from partial decodability”.

FAST against its challengers

Comparison

Five of the six papers that challenge FAST also ran it as a baseline. Each did so in its own setup, so compare the numbers within a row, not across rows. LIBERO figures are the average success rate over its four task suites.

PaperObjection to FASTUses insteadFAST against it, in that paper’s setup
OpenVLA-OFTFeb 2025 About 750 ms between action chunks is too slow for high-frequency bimanual robots. Continuous actions, parallel decoding and an L1 loss; no action tokens. LIBERO, fine-tuned: π0-FAST 85.5%, π0 94.2%, OpenVLA-OFT 97.1%.
BEASTJun 2025 Chunks of the same length give token sequences of different lengths, which blocks parallel decoding. Fixed-length B-spline tokens that need no tokenizer training. Same Florence-2 model, no robot pre-training: BEAST matches FAST on LIBERO-Long and beats it on CALVIN ABC→D. Fine-tuned π0-FAST had no success on LIBERO-Long before 30k steps.
VLA-0Oct 2025 Custom tokenizers and action heads “often require intricate training pipelines”. Actions written as plain-text numbers; the VLM is not changed. LIBERO: π0-FAST 86.0% with robot pre-training and 71.8% without; VLA-0 94.7% without.
ActionCodecFeb 2026 Its “fixed geometric priors” limit transfer across robots, and a decoded sequence must have an exact length, “a constraint frequently violated by autoregressive VLMs”, so the authors zero-pad or truncate it. A learned tokenizer designed from VLA training behaviour. Same SmolVLM2-2.2B, only the tokenizer swapped, LIBERO: FAST 90.6%, the best of the five existing tokenizers; ActionCodec 95.5%. After 5,000 training steps: 38.6% against 89.5%.
OATFeb 2026 Not every token sequence decodes, and the invalid ones make the policy “unstable, overly aggressive”. Learned, left-to-right ordered tokens, where any prefix decodes to an action chunk. Small policies trained from scratch, LIBERO: FAST 23.0%, diffusion policy 36.6%, OAT 56.3%.
The Compression GapApr 2026 A fixed-capacity discrete code caps how much a better vision encoder can help. An analysis, not a method; continuous actions avoid the cap. Not tested on FAST. Shown for OAT and diffusion policy; the author expects the same for FAST.

What the comparison shows

  • As a tokenizer, FAST holds up. When only the tokenizer changes, it beats binning and the VQ tokenizers (ActionCodec), and the newer method only matches it on LIBERO-Long (BEAST). OAT credits its low-to-high frequency order of tokens for helping autoregressive policies.
  • Decoding is the most common objection. Two papers hit token sequences that do not decode (OAT, ActionCodec), and two target its speed: the delay per chunk (OpenVLA-OFT) and the variable length that blocks parallel decoding (BEAST).
  • It needs robot pre-training. Without it, π0-FAST drops from 86.0% to 71.8% on LIBERO (VLA-0), learns late (BEAST) and does poorly in small policies trained from scratch (OAT). FAST’s own results come from π0, pre-trained on about 10k hours of robot data.
  • Its own lab answered the speed objection differently. π0.5, Knowledge Insulation and π*0.6 keep FAST tokens for training and use a flow-matching expert for inference, so the slow autoregressive decoding is not used at run time.

Sources. Each paper’s own tables and text on arXiv, read 19 October 2026: the LIBERO tables of OpenVLA-OFT, VLA-0 (Table I), ActionCodec and OAT (Table I), and BEAST’s Section 5.4 and Figure 7. The quotations are from the same papers.

All papers in the graph

Click a row to find it in the timeline

FAST’s 72 references, by how often FAST cites them
Ref.PaperYearCited in FASTCitations
141 papers that cite FAST, by citations
PaperDateCitationsInfluential

Method. Citation data from the Semantic Scholar Graph API on 19 October 2026: 781 citing papers, 115 flagged as influential (they cite FAST many times or in their method), the sentences each one uses to cite FAST, and the reference lists used to draw the links between papers. Semantic Scholar has no reference list for FAST, so the 75 references were parsed from the arXiv v1 PDF; six of them (JPEG, Huffman coding, BPE from 1994, GPT-2, MiniVLA and Lai et al.) are not matched to a Semantic Scholar record. The graph shows all 72 distinct references and 141 citing papers: the 60 most cited, the 60 most cited influential ones, and every action-tokenizer paper. The extends, validates and challenges labels are my reading of each paper. The empiricist side of this paper is in Does the universal FAST+ tokenizer transfer to cable routing?

← All paper reviews