VLA reference implementation and training toolkit

OpenPI

OpenPI is Physical Intelligence’s public implementation for selected π model releases, including base checkpoints and task adaptation examples.

Source contextPrimary specReviewed
Sources 2
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

OpenPI reviewed repository revision

Physical Intelligence · Published 2026-08-24 · Source accessed 2026-09-28

Exact tag or commit inspected through official GitHub metadata; no local benchmark execution implied.

Reviewed recordLast reviewed 2 sources ↗8 sourced fields

Role in the stack

The repository covers π0, π0-FAST and π0.5 with explicitly supported implementation paths. The existence of this toolkit does not establish public weights for every later π family release.

Source contextPrimary specReviewed
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Integration decision

Examples connect DROID, ALOHA and other robot data to model preprocessing and inference. Checkpoint, dataset transform and action normalization must match the chosen embodiment.

Source contextPrimary specReviewed
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Operational and reuse limits

The framework uses Apache-2.0 code with separate component terms, including Gemma. A base checkpoint can fail on a different robot or task distribution.

Source contextPrimary specReviewed
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Infrastructure profile

Developer
Physical Intelligence
Primary specReviewed
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Category
VLA reference implementation and training toolkit
Primary specReviewed
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Exact version / scope
Git revision 215abfb217dbac7d5f1273282331b9b1866c0479
Primary specReviewed
Sources 1
OpenPI reviewed repository revision

Physical Intelligence · Published 2026-08-24 · Source accessed 2026-09-28

Exact tag or commit inspected through official GitHub metadata; no local benchmark execution implied.

Environment
Not publicly disclosed
Access model
Public code and selected model checkpoints
Primary specReviewed
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

License
Apache-2.0 code; Gemma and other component terms separate
Primary specReviewed
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Engineering fit

Task scope
Not publicly disclosed
Compute
Not publicly disclosed
Integrations
LeRobot-format datasets; DROID and ALOHA examples
Primary specReviewed
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Pricing / access
Not publicly disclosed

Additional disclosed fields

Inputs
Robot datasets, images, language and policy configuration
Primary specReviewed
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Outputs
Fine-tuned policies and inference actions
Primary specReviewed
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

What this evidence does not establish

  • The framework uses Apache-2.0 code with separate component terms, including Gemma. A base checkpoint can fail on a different robot or task distribution.
  • The pinned release or commit identifies the review target; this registry has not run performance or compatibility tests on that software.

Relationships & deployments

Related reading is an editorial crosslink. Sourced connections describe relationships reported in the cited material. A link to a versioned profile does not establish compatibility with that version unless the connection note explicitly identifies it.

develops
Physical Intelligence → OpenPI
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Connection reviewed 2026-09-28

compatible with
OpenPI → DROID Dataset
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Connection reviewed 2026-09-28

uses
OpenPI → Physical Intelligence π0
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Connection reviewed 2026-09-28

uses
OpenPI → Physical Intelligence π0.5
Sources 1
Physical-Intelligence/openpi official repository

Physical-Intelligence · Publication date not disclosed · Source accessed 2026-09-28

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

Connection reviewed 2026-09-28

Curated comparisons

Record history & verified changes

A research review records when we checked a source. It does not mark a product launch or a new deployment.

Initial reviewed record. No subsequent field change has been recorded.

Inspect the evidence

Sources & evidence

md-openpi
Physical-Intelligence/openpi official repository

Physical-Intelligence · Repository · Publication date not disclosed · Source accessed 2026-09-28

License: Apache-2.0 code; Gemma and other component terms apply separately. Factual summary and attribution only; no upstream prose, images, weights, or dataset redistributed.

Official README inspected live. Repository code terms do not automatically cover model weights, data, or third-party assets.

md-openpi-revision
OpenPI reviewed repository revision

Physical Intelligence · Repository · Published 2026-08-24 · Source accessed 2026-09-28

Factual summary and attribution only; no upstream prose, images, weights, or dataset redistributed.

Exact tag or commit inspected through official GitHub metadata; no local benchmark execution implied.