OpenVLA Team›OpenVLA Team› offers OpenVLA

OpenVLA is a 7B parameter open-source vision-language-action model pretrained on 970k robot episodes from the Open X-Embodiment dataset. Built by fine-tuning a Prismatic-7B vision-language backbone with fused SigLIP and DINOv2 encoders and a Llama 2 action decoder, it controls multiple robot embodiments out of the box and set a new state of the art for generalist manipulation policies.

Checkpoints and the PyTorch training pipeline are fully open-source on HuggingFace, and the model adapts to new robot setups with parameter-efficient LoRA fine-tuning. It outperforms prior generalist policies such as RT-1-X and Octo on multi-robot evaluations and serves as the open baseline that generalist policy builders compare against.