// agents · openvla/openvla

openvla

openvla is an open-source ai agents project by openvla (openvla/openvla on GitHub). It has 7,123 GitHub stars and 854 forks, is released under the MIT license, and was last updated on GitHub on 23 March 2025.

Open-source 7B vision-language-action model and codebase for generalist robot manipulation.

★ 7.1k stars⑂ 854 forksMITupdated 2025-03

View on GitHub ↗

▦ visual — this project's README ships no embeddable screenshot,
so here's a placeholder. The guide below is complete regardless.

// project facts

Project facts

Repository: openvla/openvla
GitHub stars: 7,123
Forks: 854
License: MIT
Last updated on GitHub: 23 March 2025
Category: AI Agents
Listing: Full install guide

// how it works

How it works

Check your hardware first
Inference and LoRA fine-tuning need a capable NVIDIA GPU; the README's LoRA example uses a single A100 80GB, and smaller GPUs down to roughly 27GB can work with reduced batch sizes.
Install for inference or training
For inference, install the minimal requirements and load the checkpoint through Hugging Face transformers; for training, create the conda environment and install the repo editable.
Run zero-shot inference
Load openvla/openvla-7b with the AutoProcessor, pass a camera image and an instruction prompt, and read out an action with predict_action.
Fine-tune on your robot
Download a BridgeData V2 or other Open X-Embodiment dataset in RLDS format (or convert your own), then run the LoRA fine-tuning script - or the newer OFT recipe the project now recommends.

// features

Why people love it

Flagship 7B checkpoint

openvla-7b is trained on a mixture of Open X-Embodiment datasets spanning 970K trajectories, with checkpoints, configs and model cards on Hugging Face.

Minimal-dependency inference

A lightweight transformers interface (AutoProcessor and AutoModelForVision2Seq) loads the model with few dependencies, and a REST API serving script is included for robot stacks without on-device compute.

LoRA and full fine-tuning

Built-in examples cover full, partial and (quantized) LoRA fine-tuning via Hugging Face PEFT, plus full fine-tuning with PyTorch FSDP for multi-GPU nodes.

Scales from 1B to 34B

PyTorch FSDP and Flash-Attention power training across model sizes, with native support for arbitrary RLDS dataset mixtures.

FAST action tokenizer support

The project documents the FAST tokenizer, which compresses action chunks into fewer tokens for much faster inference with discrete actions.

OFT fine-tuning recipe

The 2025 OFT update offers an optimized fine-tuning recipe the team recommends for faster inference, higher success rates, multiple input images and high-frequency bimanual control.

// run it

Run it yourself

Clone the repository
Clone the OpenVLA codebase for the training, fine-tuning and serving scripts.
```
git clone https://github.com/openvla/openvla.git
cd openvla
```
Create the environment and install
Set up the pinned Python 3.10 / PyTorch 2.2 environment from the README, then install the package editable.
```
conda create -n openvla python=3.10 -y
conda activate openvla
pip install -e .
```

Load the model for inference

Install the minimal requirements and load openvla-7b in a few lines of Python.

pip install -r https://raw.githubusercontent.com/openvla/openvla/main/requirements-min.txt
python -c "from transformers import AutoProcessor; AutoProcessor.from_pretrained('openvla/openvla-7b', trust_remote_code=True)"  # see README for the full inference example

// use cases

What you can build

A robotics lab fine-tuning openvla-7b with LoRA on its own WidowX demonstrations.
Benchmarking generalist policies on the LIBERO simulation suite using the repo's documented setup.
Serving a VLA over REST so a lightweight robot can offload policy inference to a GPU machine.
Robotics students studying how vision-language models are adapted to action prediction.

// faq

Questions, answered

Who maintains OpenVLA?

The OpenVLA project team maintains the repository; it is built on the Prismatic VLMs codebase and released alongside the OpenVLA research paper.

Is OpenVLA free to use?

The code is MIT-licensed. The pretrained checkpoints are derived from Llama-2, so their use is governed by the Llama Community License rather than MIT - check its terms before commercial deployment.

Does it need a paid API key?

No. Models run locally from Hugging Face checkpoints; you need your own GPU, not any hosted API.

How does OpenVLA compare with openpi?

Both are open VLA projects: OpenVLA centers on a single 7B autoregressive model with public training code and LoRA/OFT recipes, while openpi focuses on Physical Intelligence's pi-series flow-matching models.

How do I get started?

The most-starred other AI Agents projects in the LibreAI catalogue — the realistic shortlist if openvla is not the right fit.

openclaw★ 391.6k
A personal AI that really does things — on any OS, with your data staying yours.
superpowers★ 294.9k
An agentic skills framework & software development methodology that works.
ECC★ 275.2k
Agent harness optimization system with skills, memory and security for Claude Code, Codex and more
hermes-agent★ 252.0k
The agent that grows with you — a Python framework for adaptive AI agents.
deepseek-harness★ 245.5k
DeepSeek's plugin-based agent harness where everything is a plugin.

// more in agents

Get openvla

Free and open — grab the source, star the repo, and start creating.

View on GitHub ↗← More AI Agents tools

openvla

Project facts

How it works

Why people love it

Flagship 7B checkpoint

Minimal-dependency inference

LoRA and full fine-tuning

Scales from 1B to 34B

FAST action tokenizer support

OFT fine-tuning recipe

Run it yourself

What you can build

Questions, answered

Similar tools

openpi

tada

Isaac-GR00T

Qwen-AgentWorld

Jev Ultrafast

LaVague

Alternatives in AI Agents

More in AI Agents

Get openvla