Pi0-FAST

Pi0-FAST is Physical Intelligence's autoregressive vision-language-action model, released alongside the FAST research in January 2025 and shipped in the open-source openpi release in February 2025. It pairs the pi0 VLA with Frequency-space Action Sequence Tokenization (FAST): a compression-based tokenization of robot actions using the discrete cosine transform, which trains up to 5x faster than the diffusion-based pi0 while matching its performance. The release also includes FAST+, a universal robot action tokenizer trained on 1M real robot action trajectories that works as a black-box tokenizer for diverse action spaces.

The openpi repository provides base checkpoints pretrained on 10,000+ hours of robot data for both pi0 and pi0-FAST, with examples for out-of-the-box use and fine-tuning. Pi0-FAST was developed for the team's own robot platforms, so transfer to other embodiments such as ALOHA or DROID is experimental but encouraged.