VLA-Handbook: a Chinese-language engineering manual for Vision-Language-Action work
本项目旨在为致力于进入VLA(Vision-Language-Action)领域的算法工程师提供一份全中文、实战导向的学习/面试手册。 不同于通用的 CV/NLP 面试指南，本项目聚焦于 Robotics 特有的挑战
At a glance
- What is it?
- The repository collects paper explanations, community debugging notes and GitHub issue distillations for people building robot policies. Its value is the deployment layer, not the theory index, and its licence and automation model are worth checking before you depend on it.
- Who is it for?
- Adopt VLA-Handbook if you read Chinese and your bottleneck is the gap between a paper and a working policy, particularly multi-modal synchronisation, Sim2Real breakpoints and action-space alignment. Do not adopt it as your only theory source if you need English material, and do not treat the community notes as reproducible measurements, since the README describes them as distilled from social posts and issues rather than from controlled runs.
- Can I use it commercially?
- Yes, with credit. CC-BY-4.0 allows commercial use as long as you credit the authors and indicate what you changed. It is written for creative content, so check how it applies to any code.
- Is it still maintained?
- Yes. The repository received new commits within the last day.
- What is it written in?
- Mainly HTML, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The gap this repository is trying to close
VLA papers appear in volume, and the README states the problem plainly: the engineering details that actually matter end up scattered across GitHub issues and paper appendices. Its stated goal is to fill the distance between reading a paper and getting the code to run. That is a narrower and more useful target than a general CV or NLP interview guide, because the hard parts of robot learning are not model architecture. They are multi-modal synchronisation, Sim2Real breakpoints, action-space alignment, and hardware selection for tactile sensors and dexterous hands. The intended reader is an algorithm engineer trying to enter the Vision-Language-Action field, and the material is written in Chinese throughout. If you cannot read Chinese, most of the repository is closed to you, and the README itself does not present an English edition of the main theory documents.
How the content is organised and where it comes from
The README describes 525 theory documents spread across 10 topic directories, with two to three new deep analyses added daily. Alongside that sit 165 English community field notes, more than 300 distilled Chinese community entries, 47 distilled GitHub issues from six core repositories, and 24 biweekly inference reports. The community notes are the part the README singles out as most distinctive, and they come from three sources: Xiaohongshu posts, Hugging Face blog posts plus vendor blogs and the LeRobot Discord, and GitHub issues. The issue distillation is explicitly scoped to GPU compatibility matrices covering RTX 50 series and Jetson, Pi0 fine-tuning pitfalls, GR00T memory optimisation, and root causes of training divergence. Treat these as leads rather than verified results. They are collected and distilled by a pipeline, and the README does not describe a reproduction protocol for any of the numbers that appear in them.
The reading path the repository recommends
The README lays out a dependency-ordered entry sequence rather than a flat index, which is a real editorial decision. It starts with theory/vla-core/vla_arch.md, described as a five minute overview of the path from vision and language inputs through a backbone and action head to robot actions, tracing RT-1 through RT-2, OpenVLA and pi0. The second piece, theory/diffusion-flow/action_representations.md, covers discrete tokens, diffusion and flow matching as three action-generation paradigms, and the README calls this the key fork in the field. The third, theory/diffusion-flow/pi0_flow_matching.md, explains why flow matching won, contrasting ODE straight-line paths against diffusion's curved denoising, and states 5 to 20 inference steps at 50Hz control. A fourth piece, theory/vla-core/vla_research_mainline.md, positions ACT and DP as remaining baselines and names three improvement lines: data scaling, perception enhancement and RL post-training. The ordering is the strongest signal in the README that someone thought about how a newcomer accumulates context.
Automation, feeds and the daily pipeline
The repository is not a static document set. The README describes a daily pipeline that scores new papers into grades, produces deep analyses, and writes them into the repository. PULSE.md is generated daily and tracks paper volume, acceleration and a 30-day trend chart across 15 VLA method families. Four RSS feeds are published: vla-theory.xml for new theory articles, vla-daily.xml for graded papers and the SOTA table, ai-daily.xml for agent ecosystem picks, and weekly.xml for the biweekly reports. An OPML file at rss/opml.xml imports all four into Feedly, Inoreader or NetNewsWire, and docs/SUBSCRIBE.md carries reader-specific instructions plus a CC BY 4.0 note and an FAQ. There is also a companies directory with an industry radar that the README says scans robotics and embodied AI company news (funding, products, IPOs, partnerships) daily and appends verified entries to company files, with a weekly industry judgement map. The automation is the reason the repository claims currency, and it is also the reason you should check dates on anything you cite.
Licence and what it means for reuse
The repository is licensed CC BY 4.0, confirmed by the badge in the README and the licence field on the repository. That is a content licence, not a software licence, which fits a project whose primary language is HTML and whose output is documents. Attribution is required for reuse, and the licence permits adaptation and commercial use. This is not legal advice, and the practical complication is provenance: the community field notes are distilled from Xiaohongshu posts, Discord messages, vendor blogs and GitHub issues. The repository's own licence does not settle the rights status of the underlying posts, and the README does not describe a permissions process for them. If you plan to republish the community notes rather than the theory documents, check the original sources yourself. The theory documents, being the project's own analysis, are the cleaner material to reuse under attribution.
Where it is the wrong tool
Three cases stand out. First, if you need English, the core theory corpus is Chinese and the README offers no translated edition, so the English community notes are a partial substitute at best. Second, if you need reproducible benchmark numbers, this is not that kind of project: the README's own framing is that the notes capture what papers leave out, and the GitHub issues are described as distilled from high-interaction threads, which is anecdotal by construction. Third, if you need a stable API or library to build on, there is nothing here to import. The repository is documentation plus a pipeline, and the pipeline's outputs change daily, so a link you cite today may point at a revised document next week. There are no releases retrieved for this repository, which means there is no versioned snapshot to pin to. Anyone who needs a citable, frozen reference should archive the specific file they used.
How it compares with the alternatives the README names
The README is unusually direct about competing channels and gives each one credit before contrasting. Chinese tech media such as 机器之心, 量子位 and PaperWeekly produce readable Chinese summaries with editorial quality control, and they suit mobile reading in spare moments. Awesome lists and public surveys work as bookmark collections for finding classic papers and open source projects. Following VLA authors on X or Twitter gives real-time reactions. Xiaohongshu carries first-hand failure reports and reproduction parameters that papers and media articles do not contain, though the README notes search quality is poor and posts sink quickly. The claimed difference is depth plus persistence: entry scripts, key hyperparameters and shape sanity checks, with Git history making everything greppable. That last point is the concrete distinction. A social post about a failed training run disappears; a file in a Git repository does not, and you can diff it when it changes.
Maintenance cost and what to verify before you depend on it
The last push recorded for the repository is 2026-09-10, and the README describes daily automated additions plus weekly community scans, so the maintenance burden sits with the project rather than the reader. That cuts both ways: you get freshness, but the corpus grows faster than any individual can read, and the README's own update note mentions a jump from 254 to 380 documents in a single two-month sync, with 200+ broken internal links repaired in that pass. Broken links are a recurring cost in a repository this size. If you fork it for a team wiki, budget for link checking and for deciding which of the 10 theory directories you actually care about. Before depending on any deployment advice, verify it against your own stack, starting with deployment/community_field_notes_github.md for GPU and fine-tuning specifics and deployment/README.md for the hardware selection and Sim2Real checklist. The sibling repository Spatial-Intelligence-Handbook covers world representation and 3D-aware VLA, and the README positions the two as complementary rather than overlapping.
Editorial conclusion
Adopt VLA-Handbook if you read Chinese and your bottleneck is the gap between a paper and a working policy, particularly multi-modal synchronisation, Sim2Real breakpoints and action-space alignment. Do not adopt it as your only theory source if you need English material, and do not treat the community notes as reproducible measurements, since the README describes them as distilled from social posts and issues rather than from controlled runs. Before relying on it, open theory/vla-core/vla_arch.md and deployment/community_field_notes_github.md and check whether the version-specific details (GPU compatibility, fine-tuning parameters) still match the library versions you have installed.
Community notes