VLA fine-tuning method

OpenVLA-OFT

OpenVLA-OFT is an optimization and fine-tuning recipe that changes action decoding and representation for OpenVLA, with a separately documented bimanual extension.

Source contextPrimary specReviewed 2026-09-28

Sources 3

moojink/openvla-oft official repository

moojink · 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.

OpenVLA-OFT research project

Moo Jin Kim, Chelsea Finn and Percy Liang · Publication date not disclosed · Source accessed 2026-09-28

Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success, v1

Moo Jin Kim, Chelsea Finn and Percy Liang · Published 2025-02-27 · Source accessed 2026-09-28

Table I final row and evaluation protocol inspected. Factual per-suite results only; no leaderboard copied.

Reviewed recordLast reviewed 2026-09-283 sources ↗4 sourced fields

Official source Suggest a correction

Architecture and scope

Parallel decoding, action chunking and a continuous L1 objective form the core recipe. The OFT+ extension adds feature modulation for the ALOHA setting; it is not the identical configuration behind every LIBERO row.

Source contextPrimary specReviewed 2026-09-28

Sources 3

moojink/openvla-oft official repository

moojink · 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.

OpenVLA-OFT research project

Moo Jin Kim, Chelsea Finn and Percy Liang · Publication date not disclosed · Source accessed 2026-09-28

Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success, v1

Moo Jin Kim, Chelsea Finn and Percy Liang · Published 2025-02-27 · Source accessed 2026-09-28

Table I final row and evaluation protocol inspected. Factual per-suite results only; no leaderboard copied.

Integration and evaluation

The imported LIBERO suite results use the final Table I configuration with wrist images and proprioception. Each suite is fine-tuned independently, so the results are not zero-shot generalization.

Source contextPrimary specReviewed 2026-09-28

Sources 3

moojink/openvla-oft official repository

moojink · 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.

OpenVLA-OFT research project

Moo Jin Kim, Chelsea Finn and Percy Liang · Publication date not disclosed · Source accessed 2026-09-28

Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success, v1

Moo Jin Kim, Chelsea Finn and Percy Liang · Published 2025-02-27 · Source accessed 2026-09-28

Table I final row and evaluation protocol inspected. Factual per-suite results only; no leaderboard copied.

Evidence boundary

Paper results are author-reported and were not reproduced by PhysicalAI.best.

Source contextPrimary specReviewed 2026-09-28

Sources 3

moojink/openvla-oft official repository

moojink · 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.

OpenVLA-OFT research project

Moo Jin Kim, Chelsea Finn and Percy Liang · Publication date not disclosed · Source accessed 2026-09-28

Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success, v1

Moo Jin Kim, Chelsea Finn and Percy Liang · Published 2025-02-27 · Source accessed 2026-09-28

Table I final row and evaluation protocol inspected. Factual per-suite results only; no leaderboard copied.

Model identity

Developer: Moo Jin Kim, Chelsea Finn and Percy Liang
Primary specReviewed 2026-09-28
Sources 1
moojink/openvla-oft official repository
moojink · 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: OpenVLA-OFT; arXiv:2502.19645v1 result profile
Primary specReviewed 2026-09-28
Sources 1
moojink/openvla-oft official repository
moojink · 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 fine-tuning method
Primary specReviewed 2026-09-28
Sources 1
moojink/openvla-oft official repository
moojink · 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.
Access model: Public code and task-specific checkpoints
Primary specReviewed 2026-09-28
Sources 1
moojink/openvla-oft official repository
moojink · 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: Not publicly disclosed

Inputs, outputs & embodiment

Inputs: Not publicly disclosed
Outputs: Not publicly disclosed
Embodiment: Not publicly disclosed

Implementation & training

Compute: Not publicly disclosed
Training-data disclosures: Not publicly disclosed
Integrations: Not publicly disclosed

Reported benchmark results

Context before scores. Results from different benchmarks are not directly comparable. A simulation result does not establish real-world reliability, safety or commercial availability.

Model / revision	Benchmark / metric	Reported result	Environment	Evidence & scope
OpenVLA-OFTOpenVLA-OFT, arXiv:2502.19645v1 Table I final row; wrist camera + proprioception	LIBEROTask success rate (%) · LIBERO-Spatial	97.6Reported 2025-02-27	simulationSimulated Franka Emika Panda	Primary releaseVerified 2026-09-28 Sources 1 Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success, v1 Moo Jin Kim, Chelsea Finn and Percy Liang · Published 2025-02-27 · Source accessed 2026-09-28 Table I final row and evaluation protocol inspected. Factual per-suite results only; no leaderboard copied. Methodology & comparability Per-suite fine-tuning; best checkpoint selected from periodic evaluations; 500 trials per suite. Author-reported simulation, not an independent reproduction or physical reliability estimate. Comparison group: oft-2502.19645v1-table1-wrist-proprio Paper explicitly CC BY 4.0. Attributed factual result only; no dataset or leaderboard redistributed. Verified 2026-09-28
OpenVLA-OFTOpenVLA-OFT, arXiv:2502.19645v1 Table I final row; wrist camera + proprioception	LIBEROTask success rate (%) · LIBERO-Object	98.4Reported 2025-02-27	simulationSimulated Franka Emika Panda	Primary releaseVerified 2026-09-28 Sources 1 Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success, v1 Moo Jin Kim, Chelsea Finn and Percy Liang · Published 2025-02-27 · Source accessed 2026-09-28 Table I final row and evaluation protocol inspected. Factual per-suite results only; no leaderboard copied. Methodology & comparability Per-suite fine-tuning; best checkpoint selected from periodic evaluations; 500 trials per suite. Author-reported simulation, not an independent reproduction or physical reliability estimate. Comparison group: oft-2502.19645v1-table1-wrist-proprio Paper explicitly CC BY 4.0. Attributed factual result only; no dataset or leaderboard redistributed. Verified 2026-09-28
OpenVLA-OFTOpenVLA-OFT, arXiv:2502.19645v1 Table I final row; wrist camera + proprioception	LIBEROTask success rate (%) · LIBERO-Goal	97.9Reported 2025-02-27	simulationSimulated Franka Emika Panda	Primary releaseVerified 2026-09-28 Sources 1 Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success, v1 Moo Jin Kim, Chelsea Finn and Percy Liang · Published 2025-02-27 · Source accessed 2026-09-28 Table I final row and evaluation protocol inspected. Factual per-suite results only; no leaderboard copied. Methodology & comparability Per-suite fine-tuning; best checkpoint selected from periodic evaluations; 500 trials per suite. Author-reported simulation, not an independent reproduction or physical reliability estimate. Comparison group: oft-2502.19645v1-table1-wrist-proprio Paper explicitly CC BY 4.0. Attributed factual result only; no dataset or leaderboard redistributed. Verified 2026-09-28
OpenVLA-OFTOpenVLA-OFT, arXiv:2502.19645v1 Table I final row; wrist camera + proprioception	LIBEROTask success rate (%) · LIBERO-Long	94.5Reported 2025-02-27	simulationSimulated Franka Emika Panda	Primary releaseVerified 2026-09-28 Sources 1 Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success, v1 Moo Jin Kim, Chelsea Finn and Percy Liang · Published 2025-02-27 · Source accessed 2026-09-28 Table I final row and evaluation protocol inspected. Factual per-suite results only; no leaderboard copied. Methodology & comparability Per-suite fine-tuning; best checkpoint selected from periodic evaluations; 500 trials per suite. Author-reported simulation, not an independent reproduction or physical reliability estimate. Comparison group: oft-2502.19645v1-table1-wrist-proprio Paper explicitly CC BY 4.0. Attributed factual result only; no dataset or leaderboard redistributed. Verified 2026-09-28

What this evidence does not establish

Paper results are author-reported and were not reproduced by PhysicalAI.best.
Pricing and deployment service terms were not verified.

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.

evaluated on

OpenVLA-OFT → LIBERO

Sources 1

Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success, v1

Moo Jin Kim, Chelsea Finn and Percy Liang · Published 2025-02-27 · Source accessed 2026-09-28

Table I final row and evaluation protocol inspected. Factual per-suite results only; no leaderboard copied.

Connection reviewed 2026-09-28

uses

OpenVLA-OFT → OpenVLA

Sources 1

moojink/openvla-oft official repository

moojink · 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

Record history & verified changes

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

2026-09-28

Research review / additionRegistry addition: exact paper revision, task splits, simulated embodiment and reuse terms recorded.

Four contextualized LIBERO suite results added

Sources 1

Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success, v1

Moo Jin Kim, Chelsea Finn and Percy Liang · Published 2025-02-27 · Source accessed 2026-09-28

Table I final row and evaluation protocol inspected. Factual per-suite results only; no leaderboard copied.

Inspect the evidence

Sources & evidence

md-oft

moojink/openvla-oft official repository

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

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.

Referenced by 4 records: OpenVLA · OpenVLA-OFT · What Is a VLA Model? · How to Read a Robot Benchmark

md-oft-project

OpenVLA-OFT research project

Moo Jin Kim, Chelsea Finn and Percy Liang · Paper · Publication date not disclosed · Source accessed 2026-09-28

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

Referenced by 1 record: OpenVLA-OFT

md-oft-paper

Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success, v1

Moo Jin Kim, Chelsea Finn and Percy Liang · Paper · Published 2025-02-27 · Source accessed 2026-09-28

License: CC BY 4.0. Factual summary and attribution only; no upstream prose, images, weights, or dataset redistributed.

Table I final row and evaluation protocol inspected. Factual per-suite results only; no leaderboard copied.

Referenced by 2 records: LIBERO · OpenVLA-OFT