Curiosity · AI Model
Physical Intelligence π0
π0 is Physical Intelligence's first publicly described robot foundation model. It fuses a PaliGemma-based vision-language backbone with a flow-matching action head that outputs continuous, high-frequency joint trajectories across many embodiments — two-arm, mobile manipulators, humanoids. Trained on a curated mix of open-source and in-house teleop data, π0 aims to be a general policy that can be fine-tuned (π0-FAST) or prompted for folding laundry, busing tables, and other long-horizon chores.
Model specs
- Vendor
- Physical Intelligence
- Family
- π-series
- Released
- 2024-10
- Context window
- 1 tokens
- Modalities
- text, vision, code
Strengths
- Continuous flow-matching action head — smoother than tokenised VLAs
- Multi-embodiment training — works across arms, wheels, humanoids
- Stronger dexterous performance than RT-2 on folding/cloth tasks
- Open weights released for research use
Limitations
- Commercial deployment remains Physical Intelligence exclusive
- Requires GPU-class on-robot compute for real-time inference
- Benchmarks are mostly qualitative — no standardised leaderboard
- Limited public evaluation outside PI's demonstration tasks
Use cases
- Dexterous bi-manual manipulation research
- Long-horizon household tasks — folding, busing, packing
- Cross-embodiment policy transfer (mobile-manipulator ↔ humanoid)
- Fine-tuning recipes (π0-FAST) for fleet-specific skills
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| Laundry folding end-to-end | qualitative success on novel garments | 2024-10 |
| π0-FAST fine-tune efficiency | 5× faster than diffusion-policy baselines | 2025-01 |
Frequently asked questions
What is π0?
π0 is a generalist robot foundation model from Physical Intelligence that combines a vision-language backbone with a flow-matching action head, trained on multi-embodiment robot data for dexterous manipulation.
How does π0 differ from RT-2?
RT-2 tokenises actions and emits them autoregressively. π0 uses a flow-matching continuous action head, which produces smoother high-frequency trajectories and appears to handle cloth/dexterous tasks better.
Is π0 open-source?
Weights are released for research use under a non-commercial licence. The π0-FAST fine-tuning recipe has also been published.
Sources
- Physical Intelligence — π0 blog — accessed 2026-04-20
- π0 technical report (arXiv) — accessed 2026-04-20