2024 · Stanford, UC Berkeley, Toyota Research Institute · Advanced
OpenVLA: An Open-Source Vision-Language-Action Model
OpenVLA is an open 7B-parameter VLA trained on the Open X-Embodiment dataset.
Direct answer
What does OpenVLA: An Open-Source Vision-Language-Action Model contribute?
OpenVLA is an open 7B-parameter VLA trained on the Open X-Embodiment dataset.
Background
OpenVLA combines a pretrained vision-language backbone with action prediction and is evaluated across multiple robot embodiments. The release includes model weights, training code, and fine-tuning support.
Problem
The work addresses a central constraint in VLA: building systems that learn useful representations or actions while remaining general enough to transfer beyond a single demonstration or environment.
Core idea
OpenVLA is an open 7B-parameter VLA trained on the Open X-Embodiment dataset.
Architecture and method
OpenVLA combines a pretrained vision-language backbone with action prediction and is evaluated across multiple robot embodiments. The release includes model weights, training code, and fine-tuning support.
- Open model and code
- Cross-embodiment training
- Efficient fine-tuning recipes
Results and impact
It lowered the barrier to reproducing and extending general-purpose VLA research.
Prerequisites
- RT-X
- Vision-language models
- Robot action spaces
Recommended reading order
Read the explanation above, review the related topic pages, then use the primary-source links below to inspect the abstract, figures, experiments, and released implementation.
Primary sources
External links are provided after the context needed to evaluate the work.
Follow-up research
Related papers and concepts
RT-1: Robotics Transformer for Real-World Control at Scale
RT-1 trains one transformer policy on a large multi-task dataset of real robot demonstrations.
RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control
RT-2 co-trains vision-language models on web and robot data so semantic knowledge can influence actions.
Open X-Embodiment and RT-X
Open X-Embodiment combines robot datasets across institutions and trains policies that transfer across embodiments.
Gemini Robotics 1.5
Gemini Robotics 1.5 turns visual observations and instructions into motor commands while supporting multi-step physical tasks.
Common questions
Frequently asked questions
What is the main idea of OpenVLA: An Open-Source Vision-Language-Action Model?
OpenVLA is an open 7B-parameter VLA trained on the Open X-Embodiment dataset.
Why is OpenVLA: An Open-Source Vision-Language-Action Model important?
It lowered the barrier to reproducing and extending general-purpose VLA research.
What should I learn before reading OpenVLA: An Open-Source Vision-Language-Action Model?
Recommended prerequisites are RT-X, Vision-language models, Robot action spaces.