Muted demo loop from the projectCredit: Physical Intelligence — π0 blog

Fine-tune Physical Intelligence's π0 model for your robot

Arm, Hard

openpi is Physical Intelligence's open-source release of π0, π0-FAST and π0.5 vision-language-action models, with checkpoints and fine-tuning code.

View build on GitHub ↗Share on X

Fine-tune a base checkpoint on your own LeRobot-format dataset, then serve the policy to your robot. On X this month a team taught a π model to aim a nerf gun from just 116 demos.

Hardware & cost: A large GPU for fine-tuning (see README memory requirements); any arm with a LeRobot dataset.

Difficulty: Hard.

Agent install prompt (paste into Claude Code, Cursor or Codex): "Clone https://github.com/Physical-Intelligence/openpi, install with uv per the README, and run its LIBERO or ALOHA example inference, then outline fine-tuning on my LeRobot dataset."

Source: Physical-Intelligence/openpi on GitHub — https://github.com/Physical-Intelligence/openpi (facts checked against the project README, Oct 2026; prices vary by region).

Quick answers

What do you need to fine-tune Physical Intelligence's π0 model for your robot?: A large GPU for fine-tuning (see README memory requirements); any arm with a LeRobot dataset.
How hard is it to fine-tune Physical Intelligence's π0 model for your robot?: Hard.

How your agent drives it

Copied from the Physical-Intelligence/openpi README

Shell
```
$ GIT_LFS_SKIP_SMUDGE=1 uv sync
```

Shell

$ XLA_PYTHON_CLIENT_MEM_FRACTION=0.9 uv run scripts/train.py pi05_libero --exp-name=my_experiment --overwrite

Shell

$ uv run scripts/serve_policy.py policy:checkpoint --policy.config=pi05_libero --policy.dir=checkpoints/pi05_libero/my_experiment/20000

Any coding agent that can run a shell can run these. Read the README before you run anything on a robot that moves.

Ideas to try with it

Our ideas, not the builder's claims.

Fine-tune π0.5 on your own arm demos
Serve the policy from a GPU box and stream actions to a robot over the network
Try a released checkpoint on a simulated task first

1 more idea

Compare π0-FAST and π0.5 on the same task

Source: Physical-Intelligence/openpi on GitHub
Last commit: 2026-08-24
Builder: Physical-Intelligence
Repo: Physical-Intelligence/openpi, 14,516 stars
Built with: SO-100/SO-101 arm (LeRobot), ALOHA

More like this: SO-101 / LeRobot arms · Arm manipulation

What to build next

Builds that share a board, robot or agent with this one.