Illustration brief the model wrote
A robot arm holds a cube, its trajectory shown as a wavy line. A magnifying glass reveals the line transformed into a neat stack of colored blocks, like a musical scale, indicating compression of continuous motion into discrete tokens.
FAST: Efficient Robot Action Tokenization
FAST tokenizes continuous robot actions into discrete tokens using frequency-domain compression, enabling efficient autoregressive training of vision-language-action models.
What changed
ModelBefore FAST, tokenizing continuous robot actions required high-resolution discretization or special architectures, leading to slow training and high cost. FAST introduces a frequency-domain tokenizer that compresses action chunks into a compact discrete sequence, allowing vision-language-action models to be trained autoregressively at much higher speed and lower cost without losing dexterity.
So what, in plain English
ModelImagine trying to write down every tiny movement of a robot arm in a language the robot understands. FAST turns those movements into a short code using a trick similar to how JPEG compresses pictures, so the robot can learn much faster and cheaper.
Contribution to industry
ModelThis paper contributed a practical tokenization method that reduces vision-language-action model training time and cost by roughly an order of magnitude, making fast action tokenization a standard ingredient in commercial robot models from Physical Intelligence. It enables deployment of more capable robot control policies with less computational resources.
Typed judgments
Jev★★★☆☆ robots · Physical AI 96%
Real hardware 40% · Open code / data 10% · Foundation-model based 54% · Survey / benchmark-only 3%
Probabilities from TypeSafe Jev over the title, venue and abstract; the filters on the atlas apply code-owned thresholds to these numbers.
Editor's notes
Summary. Compresses continuous robot action chunks with a frequency-domain tokenizer so vision-language-action models can be trained autoregressively at high speed without losing dexterity.
Industry. Cut VLA training time and cost by roughly an order of magnitude, making fast action tokenization a standard ingredient in commercial robot models.