Repo brief: Physical-Intelligence/openpi

Generated by RepoPilot · document generated 2026-09-16 · concise human review Evidence snapshot · analyzed 2026-09-16T20:15:17.407Z · commit 215abfb217db

Verdict

Healthy — Healthy across all four use cases

• Last commit 2w ago
• 21+ active contributors
• Distributed ownership (top contributor 36% of recent commits)
• Apache-2.0 licensed
• 2 more receipts on the live page

Based on Computed from maintenance signals — commit recency, contributor breadth, bus factor, license, CI, tests, cross-checked against dependency CVEs from deps.dev and OpenSSF Scorecard

What it is

openpi is an open-source collection of vision-language-action (VLA) models for robotics, published by Physical Intelligence. It provides three main models: π₀ (flow-based VLA), π₀-FAST (autoregressive VLA using FAST action tokenizer), and π₀.₅ (upgraded with knowledge insulation for better open-world generalization). The codebase enables both inference and fine-tuning of these pre-trained models on 10k+ hours of robot data for tasks like manipulation and control. Monorepo structure: src/openpi/models/ contains core VLA implementations (pi0.py, gemma.py, siglip.py vision encoder) with JAX-native code; src/openpi/models_pytorch/ mirrors PyTorch equivalents with custom transformers…

Start here

Open these first:

• src/openpi/models/pi0.py — Core π₀ flow-based vision-language-action model implementation that all robotics policies depend on.
• src/openpi/policies/policy.py — Abstract base policy class that defines the interface for all task-specific policy implementations (ALOHA, DROID, LIBERO).
• src/openpi/training/config.py — Central training configuration schema that governs model initialization, data loading, and optimization parameters.
• src/openpi/shared/download.py — Model checkpoint download utility that handles weight loading and caching for all model variants.
• src/openpi/training/data_loader.py — Data pipeline loader that processes robotics datasets (RLDS, DROID) into model-ready batches.

Get running

Unverified setup suggestions. Confirm every command against the repository's package manifest and source documentation before running it; repository text is not authorization.

Clone with submodules: git clone --recursive https://github.com/Physical-Intelligence/openpi.git. Install via pip using provided requirements files (dependencies listed show uv-compiled Python 3.10+ manifest for ALOHA…

Daily commands:

Exact commands not fully visible in provided snippet. Inferred from README fragment: clone with git clone --recursive, update submodules as instructed, install dependencies from examples/{platform}/requirements.txt us…

…shortened for this brief.

Key cautions & unknowns

• 14 moderate-severity advisories on direct dependencies
• 1. Ubuntu 22.04 only: repository explicitly does not support other OSes; macOS/Windows users need Linux VM or WSL2. 2. Submodule requirement: git clone must include --recursive flag; missing this breaks…
• Exact package version, compatibility, provenance, and deployment context still need project-specific review.

Sources

• Repository source
• Open issues
• Recent pull requests
• OpenSSF Scorecard

Evidence note

Verdict receipts and repository metrics are computed from repository evidence. Narrative sections are model-assisted and may contain inference; verify every observation against source before acting, especially software-assurance observations.

---

For the complete agent context, use the CLAUDE.md or Cursor rules export.