OpenVLA

7B-parameter Vision-Language-Action model for robotic manipulation

License

MIT

Overview

A 7B-parameter open-source Vision-Language-Action model for robotic manipulation. OpenVLA is trained on 970k real-world robot demonstrations and outperforms RT-2-X (55B) by 16.5% in absolute task success rate across 29 tasks, with 7x fewer parameters. Built on Llama 2 language model combined with DINOv2 and SigLIP visual encoders, it can be fine-tuned on consumer GPUs and demonstrates strong generalization in multi-task environments with robust language grounding abilities.

OpenVLA is a 7B-parameter open-source Vision-Language-Action (VLA) model designed for general-purpose robotic manipulation. Trained on 970,000 real-world robot demonstrations across 29 tasks, it achieves state-of-the-art performance while being 7x smaller than comparable models like RT-2-X (55B parameters). Built on the Llama 2 language model combined with DINOv2 and SigLIP visual encoders, OpenVLA demonstrates exceptional generalization capabilities in multi-task environments and can be fine-tuned on consumer GPUs, making advanced robotics AI accessible to researchers and developers worldwide.

Why it matters

OpenVLA demonstrated that a 7B-parameter model can outperform much larger models (RT-2-X at 55B) for robot manipulation, making vision-language-action models practical on consumer hardware. It proved that open-source robot foundation models can compete with proprietary alternatives.

Problem it solves

Previous robot foundation models like RT-2-X required enormous compute (55B parameters), making them inaccessible to most researchers. OpenVLA shows that efficient architectures can achieve better results at a fraction of the cost, democratizing robot foundation model research.

Key takeaways

• Outperforms RT-2-X (55B) by 16.5% with 7x fewer parameters
• Built on Llama 2 + DINOv2 + SigLIP — leverages proven LLM and vision backbones
• Fine-tunable with LoRA on a single GPU (the authors use one A100)
• Trained on 970K real-world robot demonstrations
• Strong language grounding enables instruction-following across 29 tasks

Who it is for

• ML researchers with limited compute budgets
• Robotics labs fine-tuning for specific tasks
• Students exploring vision-language-action models
• Companies evaluating robot foundation model approaches

Limitations

• Trained primarily on tabletop manipulation — limited to similar scenarios
• Requires GPU for inference (not suitable for edge deployment without optimization)
• Fine-tuning still requires real robot demonstration data
• Language grounding works best for simple task instructions