π0 (pi-zero)
Flow-matching vision-language-action model producing 50 Hz action chunks. Trained on 10k+ hours across 7 robot platforms and 68 tasks; folds laundry, bags groceries and assembles boxes end to end.
- 82%
- 88K
- 50 Hz
- 900K
Benchmarks
Numbers as reported by the authors on the named evaluation suite.
| Evaluation | Suite | Score |
|---|---|---|
Laundry folding (out of dryer) | PI real-world | 75% |
Zero-shot task battery | PI eval (20 tasks) | 82% |
vs OpenVLA / Octo | Matched real-robot eval | 2 |
Training
Flow-matching VLA (PaliGemma + action expert)
3.3B
Large A100/H100 cluster
45K accelerator-hours
10K h
900K
VLM pre-training then flow-matching action expert; post-training on high-quality task-specific data for dexterous behaviours.
Post-training on 1-5 h of task data beats training from scratch on 50 h.
Linked datasets, robots & papers
Resolving related models, datasets and papers…