OpenPI Robot Policy Fine-Tuning
Fine-tune and serve Physical Intelligence OpenPI robot models (pi0, pi0.5) with JAX or PyTorch
- Repository
- Orchestra-Research/AI-Research-SKILLs
- Stars
- 14,195
- Licence
- MIT
- Files
- 6 files
What it does
This skill gives an AI agent a step-by-step playbook for working with Physical Intelligence's OpenPI robot control models: cloning the repo, setting up the environment, computing normalization statistics, fine-tuning on custom robot datasets, converting checkpoints between JAX and PyTorch, starting a policy inference server, and debugging common problems like memory errors. The agent follows the documented commands and config patterns instead of guessing them.
Great for
- Robotics engineers adapting pi0 or pi0.5 models to their own ALOHA, DROID, or LIBERO-style datasets
- Researchers reproducing OpenPI fine-tuning and evaluation workflows
- Developers who need a policy inference server running locally with a WebSocket client
Skip it if
- Anyone without GPU hardware or robot/simulation experience: the compute table calls for A100/H100-class GPUs and the workflows assume familiarity with robot datasets
- Fine-tuning OpenVLA or NVIDIA Cosmos Policy: the skill itself points to other skills for those models
- Controlling a physical robot without safety review: the skill covers training and serving, not safe deployment on hardware
Try asking
- “Set up the openpi repo on a blank machine and verify the serve_policy entrypoint works”
- “Fine-tune pi05_droid_finetune on my LeRobot dataset and serve the resulting checkpoint”
- “Compute norm stats for my custom config and explain why training failed”
- “Convert a JAX checkpoint to PyTorch format for deployment”
Install in 3 steps
- 1
Download the skill
The ZIP contains the 6 original files, the MIT licence and an attribution note.
openpi.zip - 2
Add it to your agent
Claude Code (all projects)
mkdir -p ~/.claude/skills && unzip openpi.zip -d ~/.claude/skills/Claude Code / Agent SDK (this project only)
mkdir -p .claude/skills && unzip openpi.zip -d .claude/skills/Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.
- 3
Ask for the task
No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.
Get better results
- Run compute_norm_stats.py again after any change to config, dataset, or transforms; the skill treats this as mandatory.
- Review the training config and dataset repo_id before launching training, since the agent will run long GPU jobs that use --overwrite on existing run names.
- Start with a small config such as pi05_droid_finetune to validate your pipeline before attempting full-scale RLDS training.
What's inside
- references/checkpoints-and-env-map.md
- references/config-recipes.md
- references/pytorch-gotchas.md
- references/remote-client-pattern.md
- references/training-debugging.md
- SKILL.md
Redistributed unmodified from Orchestra-Research/AI-Research-SKILLs under the MIT licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.