sou350121/VLA-Handbook

VLA Handbook – A Living Knowledge Base for Vision‑Language‑Action Robotics

What it is – An open‑source, continuously‑updated repository that collects, curates, and distills the rapidly‑growing VLA (Vision‑Language‑Action) literature and engineering know‑how. It bridges the gap between reading a paper and actually running the code on a robot, with daily pipelines that fetch new papers, rate them, and generate deep‑dive notes.

Why it matters – VLA research (e.g., RT‑1, RT‑2, OpenVLA, π‑0, Figure Helix) produces dozens of papers each week, but reproducibility details are scattered across GitHub issues, community posts, and supplemental material. This handbook aggregates those details, adds sanity‑check scripts, hyper‑parameter tables, shape checks, and hardware recommendations, turning a chaotic literature stream into a searchable, version‑controlled knowledge base.

Key content

• 525 theory documents covering core VLA architecture, diffusion/flow action representations, world models, tactile perception, and frontier topics.
• 165 English and 300+ Chinese community notes harvested from Reddit/Discord, Hugging Face blogs, and Xiaohongshu, with automatic summarisation and indexing.\n- 47 distilled GitHub‑Issue experiences (GPU compatibility, memory tricks, convergence failures, etc.).
• Industry radar – daily tracking of robotics/embodied‑AI companies (funding, products, IPOs) with weekly “industry judgement maps”.
• RSS feeds & OPML for automated subscription of new theory notes, daily signal feeds, AI‑agent news, and weekly reports.
• VLA Expert Skill – a plug‑in that injects the handbook’s knowledge into Claude Code, Cursor, Codex, or OpenCode assistants.
• Automation pipeline (Pulsar) – 33 cron jobs that fetch papers, rate them (⚡/🔧/📖/❌), run sanity‑check scripts, update the repository, and self‑adjust confidence on 19 domain hypotheses.

How to use it

1. Browse the theory/ folder for deep‑dive markdowns (e.g., VLA 架构总览, Flow Matching 原理拆解). Each note includes entry‑point scripts, key hyper‑parameters, and shape‑validation snippets.
2. Deploy using the deployment/README.md – hardware selection guide, multimodal sync checklist, Sim‑2‑Real tips, and a consolidated “real‑world deployment entry”.
3. Stay updated – subscribe to the RSS feeds or the OPML file; the daily “PULSE” markdown shows traffic, acceleration, and 30‑day trends for 15 VLA method families.
4. Leverage the AI skill – copy the VLA‑expert‑skill folder into your AI‑coding assistant to get instant access to the handbook’s knowledge when you ask for paper summaries, hyper‑parameter recommendations, or deployment advice.
5. Contribute – open issues or PRs to add new paper analyses, community notes, or interview question banks (question-bank/).

Who benefits

• Researchers needing a reliable reference for VLA model internals and the latest benchmark results.
• Engineers who want reproducible scripts, hardware compatibility matrices, and Sim‑2‑Real troubleshooting tips.
• Students & job‑seekers looking for curated interview questions and industry intelligence on embodied‑AI companies.
• AI‑assistant developers who want a ready‑made VLA knowledge plug‑in.

License – Creative Commons Attribution 4.0 (CC‑BY‑4.0). Content can be shared and adapted with attribution.

Getting started – Clone the repo, read theory/README.md for the learning roadmap, then follow the deployment guide to run a VLA model on your robot. RSS feeds and the Pulsar automation run automatically; no extra setup is required unless you want to contribute.

---

All details are taken directly from the repository’s README; no external assumptions have been added.