Stanford OpenVLA

2024-06-13 · Publisher : Stanford, UC Berkeley et Toyota Research Institute · Domain : Robotics

OpenVLA, a 7-billion-parameter vision-language-action model trained on 970,000 real-world demonstrations, becomes the open reference VLA, released under the MIT license.

### Background On June 13, 2024, a team led by Stanford and UC Berkeley, together with the Toyota Research Institute, released OpenVLA. Until then, the most advanced vision-language-action (VLA) models, such as RT-2, stayed closed. OpenVLA provides a complete VLA with its weights and a fine-tuning recipe. ### Specifications - **Size**: 7 billion parameters. - **Architecture**: the Llama 2 language model, combined with a visual encoder fusing DINOv2 and SigLIP features. - **Data**: 970,000 real-world robot demonstrations. - **Results**: 16.5 points higher absolute success rate than RT-2-X (55 billion parameters) across 29 tasks, with 7x fewer parameters (scores published by the authors). - **License**: code and weights under MIT. ### Why it matters OpenVLA democratizes the VLA recipe: a web-pretrained vision-language model extended with robot action learning. It became a standard baseline for later work, including π0 and SmolVLA. ### Key takeaway With OpenVLA, any lab can fine-tune a 7B VLA for its own robot, while such models had been reserved for a handful of large companies.

Tags : VLA, Open Source, Manipulation, Llama 2