ИИ-агенты
openpi
Physical Intelligence's open vision-language-action models (pi0, pi0-FAST, pi0.5) for robot policies.
★ 14,1 тыс. звёздApache-2.0
Смотреть руководство →openvla — проект с открытым исходным кодом из категории «ИИ-агенты» от openvla (openvla/openvla на GitHub). У него 7 123 звезды на GitHub и 854 форка, он распространяется по лицензии MIT и последний раз обновлялся на GitHub 23 марта 2025.
Open-source 7B vision-language-action model and codebase for generalist robot manipulation.
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.
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.
Load openvla/openvla-7b with the AutoProcessor, pass a camera image and an instruction prompt, and read out an action with predict_action.
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.
openvla-7b is trained on a mixture of Open X-Embodiment datasets spanning 970K trajectories, with checkpoints, configs and model cards on Hugging Face.
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.
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.
PyTorch FSDP and Flash-Attention power training across model sizes, with native support for arbitrary RLDS dataset mixtures.
The project documents the FAST tokenizer, which compresses action chunks into fewer tokens for much faster inference with discrete actions.
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.
Clone the OpenVLA codebase for the training, fine-tuning and serving scripts.
git clone https://github.com/openvla/openvla.git
cd openvlaSet 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 .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 exampleThe OpenVLA project team maintains the repository; it is built on the Prismatic VLMs codebase and released alongside the OpenVLA research paper.
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.
No. Models run locally from Hugging Face checkpoints; you need your own GPU, not any hosted API.
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.
Install the minimal requirements, load openvla/openvla-7b with transformers, and run the README's inference example with a sample image and instruction.
Physical Intelligence's open vision-language-action models (pi0, pi0-FAST, pi0.5) for robot policies.
An open-source speech language model from Hume.
NVIDIA Isaac GR00T N1.7 — an open foundation model for generalist robots, ready to fine-tune and deploy.
Qwen-AgentWorld: a language world model simulating MCP, terminal, web, and OS environments for agents.
Jev Ultrafast — the fastest browser agent from browser-use: one decision-model request per action cycle drives a real browser, MIT
A Large Action Model framework for building AI web agents that browse and act for you.
Другие проекты из категории «ИИ-агенты» с наибольшим числом звёзд в каталоге LibreAI — реалистичный список альтернатив, если openvla вам не подходит.
A personal AI that really does things — on any OS, with your data staying yours.
An agentic skills framework & software development methodology that works.
Agent harness optimization system with skills, memory and security for Claude Code, Codex and more
The agent that grows with you — a Python framework for adaptive AI agents.
DeepSeek's plugin-based agent harness where everything is a plugin.
Соседние проекты из категории «ИИ-агенты» в каталоге, ближайшие по числу звёзд к openvla — продолжайте изучать полку по обе стороны от него.
Product Management skills framework built on battle-tested methods for Claude Code, Cowork, Codex, and AI agents.
A framework for building agentic apps
The Enterprise-Grade Multi-Agent Orchestration Framework. Website: https://swarms.ai
Open Multi Agent — a self-hosted TypeScript agent runtime with durable approvals and verifiable run records.
Voyager — the open-ended embodied Minecraft agent that writes its own skills and keeps learning for life.
AppAgent — Tencent's multimodal agent that learns smartphone apps by exploration, then operates them.
Бесплатно и открыто — забирайте исходный код, ставьте звезду репозиторию и начинайте создавать.