humanoidversus
Compare

Robot AI model

OpenVLA-OFT

OpenVLA-OFT fine-tunes OpenVLA with parallel decoding, action chunking and continuous-action regression; OFT+ adds FiLM conditioning for ALOHA. The authors describe a 7B VLA, with MIT code/checkpoint terms and inherited Llama 2 Community License Agreement restrictions from OpenVLA.

Records last verified Oct 13, 2026 · Send a correction

Also known as OpenVLA-OFT+

Organization
Stanford University
Type
VLA
Released / introduced
Feb 27, 2025
Public weights
Public weights
Licence / scope
MIT; Llama 2 Community License Agreement
Parameters
7B
Base model
OpenVLA 7B
Humanoid association
None reported
Inputs
vision, language, proprioception
Outputs
continuous robot action chunks

Release dates can mark an announcement or paper introduction. Public checkpoints, code licences and weight terms are distinct; follow the publisher’s links for the stated scope.

Paper ↗Code ↗Weights ↗Project ↗

Documented embodiments

Source-associated hardware; this list does not establish plug-and-play compatibility or commercial robot capability.

  • ALOHA

    Real-robot bimanual evaluation uses the OFT+ variant.

  • Franka robot arm (LIBERO simulation)

    LIBERO-Spatial/Goal/Object/Long checkpoint variants.

Disclosed training data

More robot AI models

Helix →Gemini Robotics-ER →

Evidence

Sources

  1. [s1] Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success ↗

    Stanford University · paper · published Feb 27, 2025 · accessed Oct 13, 2026

  2. [s2] OpenVLA-OFT project: 7B VLA, OFT+ and ALOHA evaluations ↗

    Stanford University · official · accessed Oct 13, 2026

  3. [s3] OpenVLA-OFT official code and pretrained-checkpoint instructions ↗

    Stanford University · official · accessed Oct 13, 2026

  4. [s4] OpenVLA-OFT LIBERO-Spatial model card and MIT tag ↗

    Moo Jin Kim · official · accessed Oct 13, 2026

  5. [s5] OpenVLA repository: inherited Llama 2 restrictions and OXE pretraining ↗

    OpenVLA Collaboration · official · accessed Oct 13, 2026