VLA / foundation model
π0 / π0.5 (Physical Intelligence)
A vision-language-action model that outputs smooth 50 Hz action chunks and was pre-trained on 10,000+ hours of data from seven robot setups.
How it works
π0 bolts an action expert onto a vision-language model. Instead of emitting actions as text tokens it uses flow matching (a cousin of diffusion) to produce a chunk of continuous actions at 50 Hz, which is fast enough for dexterous tasks like folding laundry. Pre-training mixes the company's own data from single-arm, bimanual and mobile robots with open datasets. π0.5 adds broader web and multi-environment data to generalize to new homes. Both are released as open weights in the openpi repository.
What it needs
Fine-tuning on your own task: tens to hundreds of episodes collected on your setup; a capable GPU. The released checkpoints assume specific embodiments (e.g. UR5e, Trossen, ARX); other robots need adaptation.
Strengths
- Smooth high-rate actions suit dexterous and bimanual tasks
- Open weights (openpi) and active community
- Demonstrated long-horizon tasks (laundry, bussing)
Limits
- Results are reported as scores in figures, hard to compare
- Industrial pilots (Siemens) found reliability below production needs
- Large model; inference hardware is not trivial
Where it has been shown to work (2)
| Task | Robot | Maturity | Success | n | Grade | Year |
|---|---|---|---|---|---|---|
| Packaging and bagging | UR7e | Pilot | — | — | C | 2026 |
| Cable and deformable handling | ARX-5 arm | Lab | — | 10 | B | 2024 |