π0-FAST
Replaces per-dimension binning with a DCT-based action tokenizer, cutting token count 5x and making autoregressive VLAs viable at 50 Hz control rates.
- 79%
- 43K
- 50 Hz
- 850K
Benchmarks
Numbers as reported by the authors on the named evaluation suite.
| Evaluation | Suite | Score |
|---|---|---|
Training-time speedup | FAST paper | 5 |
DROID zero-shot | DROID eval | 79% |
Parity with π0 diffusion | 7 dexterous tasks | 98% |
Training
Autoregressive VLA with FAST action tokenizer
3B
H100 cluster
30K accelerator-hours
9.5K h
850K
Discrete cosine transform + byte-pair encoding over action chunks; trained autoregressively like a language model.
Tokenizer is embodiment-agnostic — reusable without retraining for new arms.
Linked datasets, robots & papers
Resolving related models, datasets and papers…