EEmbodied AI Hub
ModelFeaturedFully open

OpenVLA

Open 7B Vision-Language-Action model pretrained on Open X-Embodiment (~970k demos). Canonical paper is arXiv:2406.09246 (2024); for 2025 finetuning SOTA see OpenVLA-OFT (arXiv:2502.19645).

Why: One of the most followed open VLAs with relatively complete docs and weights

Official / sourcePaperHugging FaceBack to Model
VLAOpen sourceCross-embodiment

Overview

OpenVLA is a fully open Vision-Language-Action model trained on Open X-Embodiment. It is a landmark release because weights, code, and data recipes are available enough for outside labs to finetune and study VLA behavior without a closed API.

The Paper button points to the canonical model paper (arXiv:2406.09246, June 2024, later revised on arXiv). That date is expected: it is the definition paper for the base 7B OpenVLA checkpoint—not a sign that the ecosystem stopped. The main 2025 follow-up from the same line is OpenVLA-OFT (arXiv:2502.19645, RSS 2025), an optimized fine-tuning recipe that greatly improves LIBERO success and real-robot throughput; use the separate OpenVLA-OFT resource for that paper and code.

Architecturally OpenVLA connects a VLM backbone to action prediction for manipulation across embodiments in OXE. Practically, 7B-class models need serious GPUs for finetuning; inference benefits from chunking/OFT-style recipes.

Who it is for

VLA researchers; labs finetuning language-conditioned manipulators; open-source reproducibility work.

Key highlights

  • Fully open weights + training lineage on OXE
  • Standard open VLA baseline in the literature
  • Cross-embodiment pretraining signal
  • Ecosystem finetune recipes (e.g. OFT)

When to use

  • You need an open VLA checkpoint to finetune
  • You study language-conditioned manipulation
  • You compare against published VLA numbers

When not to use

  • Single small GPU with no finetune budget
  • Tasks far outside OXE distribution without adaptation data

Getting started

  1. 1Download weights from the official HF card; verify license.
  2. 2Run zero-shot or few-shot evals on a known benchmark (e.g. LIBERO).
  3. 3Finetune with OFT or official scripts on a small real dataset.
  4. 4Measure real success rate and latency on your robot.

Install / setup

See OpenVLA repo README

git clone https://github.com/openvla/openvla.git

Papers & reading

How it compares

Strong open VLA baseline; heavier than SmolVLA, more open than many proprietary stacks.

Caveats & pitfalls

  • OXE mixture details affect reproducibility — document your subset.
  • 7B finetunes are expensive; prefer parameter-efficient recipes when possible.

Content reviewed 2026-07-23