OpenVLA

**OpenVLA** is the first fully open-source replication of the RT-2 architecture, announced in June 2024 (paper 'OpenVLA: An Open-Source Vision-Language-Action Model', Kim et al., arXiv:2406.09246). It was developed jointly by Stanford AI Lab, UC Berkeley (Robot Learning Lab), Google DeepMind, Toyota Research Institute, MIT and Physical Intelligence. OpenVLA fills the gap left by the closed RT-2 — releasing the **model weights**, **training code**, **fine-tuning recipes**, and **complete data pipeline**.

Architecture: ~**7B parameters** built from three components. (1) **Vision encoder** — a fusion of DINOv2 (semantic features) + SigLIP (CLIP-style alignment), both ViT-L/14. (2) **LLM backbone** — Llama 2 7B. (3) **Action head** — action discretization into 256 bins per dimension (as in RT-2), next-token prediction over action tokens.

Training data: ~970,000 demonstrations from **Open X-Embodiment** (Google DeepMind, 21 institutions), covering 22 robots (Franka, UR5, WidowX, Sawyer, Google Robot etc.) and ~500 tasks. Training time: 8 days on 64× A100 80 GB.

Results: OpenVLA achieves **+16.5 pp success rate** over RT-2-X (55B) on out-of-distribution generalization tasks — despite having 8× fewer parameters. Fine-tuning on custom datasets (LoRA-style) takes 10-20 hours on 1× A100 and adapts the model to a new robot with 100-500 demonstrations.

Ecosystem: full integration with **HuggingFace Transformers** (`openvla/openvla-7b`), support for 4-bit quantization (bitsandbytes), compatibility with PyTorch 2.0+. Impact: OpenVLA has become the **de facto VLA baseline** in academia — the basis of all subsequent works (CogACT, TraceVLA, RoboFlamingo). Reproducibility: full checkpoints, dataset indices, and training scripts.