OpenVLA: An Open-Source Vision-Language-Action Model

1 min read

OpenVLA is an open 7-billion-parameter vision-language-action model trained on 970,000 real-world robot demonstrations. It can be fine-tuned on consumer GPUs with low-rank adaptation and served quantised without losing task success.

Findings

  • Beats the closed RT-2-X (55B parameters) by 16.5% absolute task success over 29 tasks, with 7 times fewer parameters.
  • Fine-tuned on new settings, beats Diffusion Policy by 20.4%.

Limits

  • Abstract-level reading only.

Relevance

The executor inside Agentic Robot, which wraps it with a planner and a verifier. Open weights make it possible to probe how a VLA fails, which closed models like RT-2 do not allow. See foundation-models-for-robotics.