UAV World Model & VLA & VLM Learning¶

World Model / VLA / VLM Learning Library — for UAVs and embodied AI

What this is • What is different • Quick start • Contents • Cite • Contributing

The documentation is written in Chinese. This file and docs/en/index.md are the English front door: they map the repository and link to the Chinese originals. The 59 documents themselves are not translated.

What this is¶

59 documents · 195 papers · 24 CPU-only demos · ~287k Chinese characters.

World models, VLAs and VLMs are the three foundations of embodied AI right now. This repository turns them into a path you can walk from the beginning: the spine is general method and literature, and aerial robotics is a running thread through all nine sections rather than a side chapter.

Choosing drones as the specialisation narrows nothing — flight is the harshest embodiment there is (4-DoF control, a crash you cannot undo, an onboard compute budget measured in watts). Where a general method breaks in the air is exactly where its limits become visible.

Main references:

arXiv:2605.00080 — World Model for Robot Learning: A Comprehensive Survey (43 pages, 2026).
NTUMARS/Awesome-World-Model-for-Robotics-Policy — a collection of world-model papers for robot policy learning; our annotations are in references/awesome-annotations.md.

The three core concepts:

VLM (vision-language model)   VLA (vision-language-action)   World Model
"describe what it sees"       "describe, then act"           "predict what comes next"
┌─────────────┐               ┌─────────────┐                ┌─────────────┐
│  image+text │               │ img+text+act│                │ state + act │
│      ↓      │               │      ↓      │                │      ↓      │
│  understand │               │ understand  │                │ predict next│
└─────────────┘               └─────────────┘                └─────────────┘
 "obstacle ahead"            "turn left to avoid"           "what if I turn left"

What is different¶

Paper lists are not scarce. This repository tries to add three things they usually lack: you can run it, track it, and test yourself on it.

Asset	What it is
24 CPU-only demos	No training, no VRAM, each produces a figure in about two minutes. Every demo verifies exactly one claim in the text — and where a claim cannot be measured, the text says so instead of implying otherwise. See `code/`.
Frontier logs for all three topics	`tools/watch.py` re-queries world models / VLA / VLM monthly, producing `references/*-watch-2026-10.md`.
Self-assessment sets	`docs/09-专题自测与考察/` — questions only; the body text gives pointers, never answers.
A hardware limit stated up front	Every number comes from a real run. The ceiling is a single 8 GB GPU, so this repository does not train large models — it covers what actually runs.

A claim in the text looks like this — every figure below comes from a rerunnable script under code/:

_{Seven navigation metrics, seven different rankings}

_{Feed the model its own output: step 60 is 1084× the error of step 1}

_{VRAM budget: three exact calculations, one measurement}

Quick start¶

git clone https://github.com/Qxy661/UAV-WM-VLA-Learning.git
cd UAV-WM-VLA-Learning
py -3.9 code/a_worldmodel_rssm.py     # CPU only, figure lands in figures/ in seconds

On Windows use py -3.9; on macOS / Linux use python3. All 24 demos need only numpy and matplotlib — no GPU, no downloads, no model weights.

If you would rather only read: start with the study route (Chinese) or the English documentation overview.

Read it online: https://qxy661.github.io/UAV-WM-VLA-Learning/ — full-text search with Chinese word segmentation, light/dark themes, rendered diagrams. GitHub Pages can be slow to reach from mainland China; if it does not load, clone the repository and serve it locally: pip install "mkdocs-material==9.7.*" "mkdocs<2" jieba, then py -3.9 tools/build_docs.py && mkdocs serve.

Contents¶

Nine sections, 59 documents. The Chinese title of each section is the directory name.

#	Section	Docs	Start
00	Study route and prerequisites	1	导读与学习路线
01	Foundations — what a world model / VLM / VLA is	5	什么是世界模型
02	World models — history, generative, model-based RL, 3D, long-horizon	9	世界模型发展史
03	VLA — architecture, aerial models, action heads, data, RL post-training	10	VLA 架构演进
04	VLMs — remote sensing, aerial agents, edge deployment, evaluation	7	遥感 VLM
05	Close reading of one survey, section by section	5	综述概览与结构
06	Paper cards — 45 papers, one card each	4	世界模型论文卡片
07	Hands-on guides — environment, 10 reproduction guides	11	环境搭建
08	Research frontier — roadmap, open problems, critical reading	4	研究路线图
09	Self-assessment — VLA / world model / VLM	3	VLA 专题自测

Also in the repository:

references/paper-list.md — all 195 papers with a ★★★ / ★★☆ / ★☆☆ recommendation scale.
mindmaps/ — 4 mind maps (field overview, world-model taxonomy, VLA timeline, reading order).
code/ — the 24 demos.
tools/ — frontier watching, link and nav linting.

Cite¶

CITATION.cff is in the repository root, so GitHub renders a Cite this repository button with BibTeX and APA exports:

@software{qxy661_uav_wm_vla_2026,
  author  = {Qxy661},
  title   = {UAV World Model \& VLA \& VLM Learning},
  year    = {2026},
  version = {1.0.0},
  url     = {https://github.com/Qxy661/UAV-WM-VLA-Learning},
  license = {MIT}
}

Contributing¶

The contribution guide is in Chinese: CONTRIBUTING.md. In short — open an issue with the file and line number, or send a PR that passes tools/check_links.py and tools/check_nav.py. Issues are welcome in Chinese or English.

License¶

MIT.