Physical Intelligence π0-FAST

2025-01-16 · Publisher : Physical Intelligence · Domain : Robotics

Physical Intelligence releases FAST, a robot action tokenization based on the discrete cosine transform, and π0-FAST, an autoregressive π0 variant up to 5x faster to train.

### Background On January 16, 2025, Physical Intelligence released FAST (Frequency-space Action Sequence Tokenization). Autoregressive VLAs must split continuous robot actions into "tokens", as a language model splits text. Naive per-timestep, per-joint binning fails on dexterous, high-frequency motion. FAST solves this with compression. ### Specifications - **Method**: a discrete cosine transform, similar to JPEG compression, applied to action sequences. - **FAST+**: a universal tokenizer trained on 1 million real robot trajectories, usable as-is across action spaces and control frequencies. - **π0-FAST**: combined with π0, FAST enables training on 10,000 hours of robot data and matches diffusion VLAs while cutting training time by up to 5x (results published by the company). - **License**: tokenizer released on Hugging Face under Apache 2.0. ### Why it matters FAST makes autoregressive VLAs, easier to train with large language model tooling, competitive on fine-grained tasks. It broadens architectural choices for robot models. ### Key takeaway π0-FAST joined the open openpi repository in February 2025, alongside π0.

Tags : VLA, Tokenisation, Open Source, Physical Intelligence