Pi
Table of contents
The π (Pi) models are a family of vision-language-action models developed by Physical Intelligence and released through the open-source openpi repository. The repo provides base checkpoints pretrained on more than 10k hours of robot data, together with training and fine-tuning code in both JAX and PyTorch.
▪ Model Family
- π₀: a flow-based VLA. A pretrained VLM backbone (PaliGemma) encodes images and language, and a smaller action expert generates continuous action chunks with flow matching, enabling high-frequency (up to 50 Hz) dexterous control.
- π₀-FAST: an autoregressive variant built on the FAST action tokenizer, which compresses action chunks into a short sequence of discrete tokens for language-model-style generation.
- π₀.₅: an upgraded version of π₀ trained with knowledge insulation, giving substantially better open-world generalization; the repo currently supports it with the flow-matching action head.
▪ Checkpoints
Base checkpoints (pi0_base, pi0_fast_base, pi05_base) are intended for fine-tuning, while “expert” checkpoints fine-tuned on specific platforms — DROID, ALOHA, and LIBERO — can often run zero-shot on those setups. Checkpoints are downloaded automatically from gs://openpi-assets and cached locally.
▪ Usage
Running a pretrained policy takes a few lines:
from openpi.training import config as _config
from openpi.policies import policy_config
from openpi.shared import download
config = _config.get_config("pi05_droid")
checkpoint_dir = download.maybe_download("gs://openpi-assets/checkpoints/pi05_droid")
policy = policy_config.create_trained_policy(config, checkpoint_dir)
action_chunk = policy.infer(example)["actions"]
Fine-tuning uses LeRobot-format datasets: convert the data, compute normalization statistics with scripts/compute_norm_stats.py, train with scripts/train.py, and serve the resulting checkpoint through a websocket policy server (scripts/serve_policy.py) so the robot client can stream observations and receive action chunks remotely.
Reference: