OpenVLA-7B

Sobre Stanford / UC Berkeley / Toyota Research Institute

Fundado em 2024 · Stanford & Berkeley, California, USA

OpenVLA is the result of an academic-industry consortium led by Moo Jin Kim and colleagues at Stanford, UC Berkeley, and Toyota Research Institute (with contributors from MIT, Google DeepMind and Physical Intelligence). Released in June 2024, it was the first fully open-weights 7-billion-parameter Vision-Language-Action model trained on the Open-X-Embodiment dataset. OpenVLA was designed as a direct, reproducible, and parameter-efficient alternative to Google's closed RT-2 / RT-2-X, with the explicit goal of letting any lab fine-tune a 7B-class VLA on a single A100 / H100. The model, code, training recipe and fine-tuning toolkits (including LoRA) are all released under MIT-style permissive licences. OpenVLA quickly became a standard baseline in academic VLA research and the starting point for many downstream policies (CogACT, π-0-FAST baselines, embodied agent demos).

Arquitetura

Vision-Language-Action (autoregressive discrete-token VLA)

OpenVLA combines a Llama-2-7B language backbone with a dual visual encoder that concatenates DINOv2 and SigLIP features, fused into the LLM via a Prismatic VLM-style projector. The model treats actions as discretised tokens: each continuous robot action dimension is binned into 256 bins, and the resulting tokens are appended to the LLM's vocabulary. Training is a single autoregressive next-token objective predicting both language and action tokens given image observations and a natural-language instruction. OpenVLA was trained on ~970k demonstration episodes from the Open-X-Embodiment dataset spanning 22+ robot embodiments and used Llama-2-7B as a pretrained text+code backbone, which the authors found markedly improves language grounding compared to scratch-trained VLAs. Parameter-efficient fine-tuning with LoRA is officially supported, making OpenVLA the de-facto open VLA workhorse.

Parâmetros

7B

Capacidades

• Fully open-weights 7B Vision-Language-Action model
• Llama-2-7B backbone with DINOv2 + SigLIP vision
• Discrete action-token decoding (256 bins per DoF)
• Trained on ~970k Open-X-Embodiment episodes
• LoRA fine-tuning officially supported
• Strong language-grounded manipulation across robots
• Fits on a single A100 / H100 with quantisation
• MIT-style permissive licence on weights and code
• Best for: research, reproducible VLA baselines, fine-tuning on new robots.

Treinamento & licença

~970,000 robot demonstration episodes from the Open-X-Embodiment dataset (RT-X collection), spanning 22+ robot embodiments and a wide range of manipulation tasks. Llama-2-7B and the DINOv2 + SigLIP vision encoders provide web-scale pretraining priors.

Licença: MIT-style permissive licence on code and weights; Llama-2 components subject to Meta's Llama-2 Community Licence. Considered research-friendly open-weights.

Testes de segurança: OpenVLA is a research artifact - no formal red-teaming. Safety in deployment is left to downstream systems (force limits, workspace constraints, e-stops).