Open Source RoboticsRobots, software & learning

OpenPI

Software

Vision-language-action model code, checkpoints and adaptation examples for pi0, pi0-FAST and pi0.5, with backend-specific features and robot-specific action conventions. [1]

Installation & first result →Official websiteGetting started

For
Robot-learning researchers able to validate model, data and robot-interface compatibility [1]
Source rights
Some licenses are unresolved. Review licenses
Unknowns
compute requirements not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
License
Apache-2.0 · license not fully assessed for model weights, datasets, documentation · Details by part

Requirements · Research details and dates

On this page

Overview

openpi provides code, published checkpoint paths and examples for Physical Intelligence's pi0, pi0-FAST and pi0.5 robot-policy models. The reviewed commit supports the flow-matching head for pi0.5. Upstream describes the release as experimental and warns that adapting a policy to another robot may not succeed. This entry treats those checkpoints as starting points for an experiment, not ready-made controllers for arbitrary hardware. [1] [2]

Is this for you?

Workflows
  • Fine-tune a base model on converted data
  • Compute or deliberately reuse normalization statistics
  • Serve inference on a separate GPU host
  • Map image/state observations to robot-specific action chunks

[1] [5] [6]

Known limitations
  • PyTorch lacks pi0-FAST, LoRA, FSDP, mixed-precision training and EMA training weights in this snapshot
  • pi0.5 implementation supports the flow-matching head
  • Fine-tuned checkpoints may not generalize to a different robot setup
  • Checkpoint and complete pretraining-data rights are not established by the code license

Rights scope is an editorial limitation; backend support follows the pinned README.

[1] [3] [8]

What you need

Operating system
  • Ubuntu 22.04

Upstream states tested and supported platform for this snapshot.

[1]

Languages / runtimes
  • Python

[1] [7]

Dependencies
  • uv-managed Python environment and repository submodules
  • JAX or supported PyTorch model implementation
  • Matching checkpoint, normalization statistics and policy configuration
  • Robot-specific observation/action transformations

[1] [5] [6]

Host computer

NVIDIA GPU; upstream single-GPU estimates exceed 8 GB for inference, 22.5 GB for LoRA fine-tuning and 70 GB for full fine-tuning.

Upstream estimates, not measurements here; LoRA support is backend-dependent.

[1]

Installation & first result

Open the upstream getting-started guide [1] · Project documentation [4] [6]

Installation

Follow the recursive-submodule and uv setup instructions, choose a backend and matched checkpoint, then validate the policy input/output mapping using the provided examples.

Editorial route; no installation or execution performed.

[1] [6] · Editorial assessment

Hardware, models & data

Interfaces
  • Policy inference API returning action chunks
  • WebSocket policy server and openpi-client
  • LeRobot training-data conversion examples

[1] [6]

Inputs and outputs

Policy-specific image/state observations and a prompt produce action chunks; joint position versus velocity, gripper convention and control frequency depend on the robot/data route.

[5] [6]

Project notes

Backend choice changes the available workflow

The pinned README supports JAX and selected PyTorch models. Its PyTorch section excludes pi0-FAST, LoRA, FSDP, mixed-precision training and EMA weights during training. Therefore, a memory estimate or fine-tuning recipe elsewhere in the overview is not automatically available through both implementations. The PyTorch setup also patches transformers 4.53.2; upstream warns that the default uv hardlink mode can propagate those modifications through the shared cache. Review that setup in an isolated environment before reusing it across projects. [1]

The requirements section says the current training script lacks multi-node support, while the later PyTorch section supplies a multi-node launch example. That discrepancy needs a backend-specific reading and validation; this profile does not assert one universal multi-node capability. Upstream's stated supported system is Ubuntu 22.04 with an NVIDIA GPU. Its single-GPU estimates are greater than 8 GB for inference, 22.5 GB for LoRA and 70 GB for full fine-tuning, not measured requirements from this review. [1]

A checkpoint needs its configuration and statistics

The README lists base and task-specific checkpoint paths in the public openpi-assets bucket. They are downloaded and cached by the example utilities when used. Selecting a path is only part of setup: the policy configuration supplies the model and data transformations, and normalization statistics accompany the checkpoint. This review inspected these instructions without downloading model payloads. [1] [5]

The normalization guide distinguishes recomputing statistics for new data from reusing statistics for a matching pretraining robot. Reuse requires the same action-space definitions, not merely the same robot brand. It recommends evaluating both choices when adapting a task. The guide's base-model conventions include joint angles in radians and gripper positions from zero for open to one for closed, but DROID uses joint-velocity actions and a 15 Hz control frequency. Those differences must be resolved before sending a predicted action to a robot. [5]

The listed normalization asset identifiers are specific to the documented base-model routes. Do not infer that every statistic applies unchanged to every newer checkpoint. The README's data-conversion workflow also computes statistics before fine-tuning and describes input/output transformations for the selected environment. [1] [5]

Remote inference separates environments, not robot semantics

The remote-inference example runs the policy on a GPU server and uses a lightweight WebSocket client in the robot environment. Its client resizes images, converts them to unsigned 8-bit format and sends policy-specific observation keys, state and a task prompt. It returns an action chunk with horizon and action dimensions. The robot-side integration still has to interpret and execute those actions according to the selected policy's conventions. [6]

Use the repository's no-robot inference example and the relevant robot or simulator instructions as staged integration checks. That is an editorial starting sequence; this publication has not run inference, training, networking or hardware control and does not report latency or task-success measurements. [1] [6]

What we checked

Source research only; no installation, physical build or performance test performed. No affiliate links added.

Not yet established

  • A step-by-step setup summary
  • Version-specific hardware compatibility (ecosystem tags alone do not establish it)

Updates

Upstream last updated
Aug 24, 2026 [2]
Latest release
Not established
Profile updated
Oct 4, 2026
Sources last checked
Oct 1, 2026
Next review
Oct 31, 2026
Reviewed revision
215abfb217dbac7d5f1273282331b9b1866c0479 (commit snapshot; no published stable release found) · software repository

Source files and licenses

Available project sources support research coverage; complete build and commercial qualification remain separate.

LayerAvailability / license
softwareavailable · Apache-2.0
Code license; repository also includes separate Gemma terms, and dependencies retain their own terms. [3]
model weightsavailable · License not fully assessed
README publishes gs://openpi-assets checkpoint paths. Payloads not downloaded. Complete checkpoint license mapping unresolved; bundled Gemma terms are not Apache-2.0 and include use restrictions. [1] [8]
datasetsavailable · License not fully assessed
Public fine-tuning data links and conversion examples exist. Full 10k+ hour pretraining mixture availability and constituent rights not established; includes references to internal data. [1]
documentationavailable · License not fully assessed
Documentation source available; separate documentation-wide rights not established. [4] [5] [6]

Separate source, checkpoint and training-data rights

The source root license is Apache-2.0. The repository also includes Gemma Terms of Use, which contain use restrictions and distribution obligations. The inspected materials do not establish a complete license mapping for every published checkpoint, so this profile does not label all weights Apache-2.0 or uniformly open source. [3] [8]

Likewise, public fine-tuning examples and dataset links do not establish release of the entire pretraining mixture. The overview describes more than 10,000 hours of training data and references both public and internal task data. Availability and rights for the complete mixture remain unresolved here. Source availability, downloadable checkpoints and full training reproducibility are distinct findings. [1]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Project sources

Videos

  • S009 Physical Intelligence (YouTube channel)
    The project’s own YouTube channel, found by searching YouTube and checked against the channel page. Not reviewed by us for accuracy.
    Revision, date & where cited

    YouTube channel Physical Intelligence · youtube.com · Accessed Oct 4, 2026

Correction guidance · Discussion is not enabled · Profile format 0.6.0-draft.1

Change history

2026-10-04 — Added 1 YouTube reference(s) (S009): official channel or videos linked from the pinned upstream sources. Existing claims, source-check dates and evidence are unchanged.