OpenVLA Released: New 7B Parameter Open-Source VLA Model Outperforms RT-2-X and Octo in Robotics

Jun 14, 2024, 04:03 PM

OpenVLA, a new open-source vision-language-action (VLA) model, has been released. Developed from Llama-2 and incorporating Dino features, OpenVLA boasts 7 billion parameters and is trained on 970,000 robot episodes from the Open X-Embodiment dataset. It outperforms existing models like RT-2-X and Octo in zero-shot evaluations while being nearly 10 times smaller than RT-2-X. The model is designed for efficient inference and fine-tuning on a single GPU, utilizing quantization and LoRA. OpenVLA's code, data, and weights are fully available online, including a PyTorch codebase and models on HuggingFace, making it a significant step forward in accessible large-scale robotic learning. The project is expected to drive advancements in both academic and industry settings.