Skip to content

OpenVLA

Control multiple robots from one open vision-language-action model

Visit official site

What OpenVLA is

The model is described as an open-source vision-language-action model pre-trained on 970k robot episodes from the Open X-Embodiment dataset, with checkpoints and a PyTorch training pipeline released openly.

What you can do with it

  • Control a Franka Panda robot for tabletop object placement
  • Deploy on a WidowX robot for object manipulation and environment interaction
  • Apply natural-language instructions on a Google robot for object handling
  • Control several robot platforms out of the box with no extra training

Getting started

  • Download the OpenVLA checkpoints from HuggingFace
  • Set up a PyTorch training environment with all dependencies
  • Fine-tune OpenVLA for the target robot platform and task
  • Use LoRA or another parameter-efficient method to refine the model

Model notes

  • Pre-training uses 970k robot episodes from the Open X-Embodiment dataset
  • Prismatic-7B VLM combines fused vision encoders, a projector and Llama 2 7B
  • Only 1.4% of the parameters are updated through LoRA fine-tuning
  • Generalisation is listed across visual, motion, physical and semantic tasks

Pros & cons

✓ What we like

  • Controls multiple robots out of the box with no extra training
  • Checkpoints and the PyTorch training pipeline are fully open source
  • Adapts to new robot setups through parameter-efficient fine-tuning

! What to watch out for

  • Fine-tuning is described for new robot platforms and task needs
  • Setup requires a PyTorch training environment with all dependencies

FAQ

How much of the model is fine-tuned with LoRA?

The material states that only 1.4% of the parameters are fine-tuned.

Where do the checkpoints come from?

The checkpoints can be downloaded from HuggingFace, and the PyTorch training pipeline is open source.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review