Documentation Overview (English)

The body of this repository is written in Chinese. This page is a map, not a translation: every document gets an English title and a one-line description, and each entry links to the Chinese original. If a title here interests you, open the link and read it with a translator — the figures, tables and code blocks carry most of the argument anyway.

• 59 documents across nine sections, plus 4 mind maps and 6 reference files.
• 195 papers, catalogued by topic with arXiv IDs and recommendation ratings.
• 24 CPU-only demos under code/ — no GPU, no training, each runnable in about two minutes.
• Drones are a running thread, not the whole scope. The spine is general world-model / VLA / VLM method and literature; aerial robotics is the specialisation running through all nine sections.

中文首页 | English README

---

Start here

• 导读与学习路线 — Reading paths, prerequisites, and the seven-stage route from background to research frontier.

1. Foundations

• What Is a World Model — Definition, history, and the four schools that all call themselves "world model".
• What Is a VLM — Vision-language model architecture and training in one pass.
• What Is a VLA — How VLA grew out of VLM by adding an action head and a control loop.
• VLM vs. VLA vs. World Model — Where the three overlap, where they genuinely differ, and why people conflate them.
• Drones vs. Ground Robots — What an aerial embodiment changes: 6-DoF, tight mass and power budgets, no safe stop, scarce data.

2. World Models

• A History of World Models — From Ha & Schmidhuber through to modern video world models.
• Generative World Models — Video prediction and scene generation for aerial views (ANWM, AirScape, FlightDiffusion).
• Model-Based RL World Models — The Dreamer line, and what it takes to fly with it.
• 3D Scene World Models — NeRF and 3D Gaussian Splatting read as world models rather than as rendering tricks.
• Aerial World Models: Overview — The papers that are specifically about UAV world models, perception to decision.
• Key Datasets and Benchmarks — MotionScape, AeroVerse, and what each one can and cannot measure.
• Evaluating and Diagnosing World Models — One-step accuracy is not rollout usability; the diagnostics that separate them.
• JEPA and Latent World Models — Predicting in representation space instead of reconstructing pixels, and what that buys.
• Long-Horizon and Interactive Generation — Drift, forgetting, and lost object permanence once the horizon passes a hundred steps.

3. VLA (Vision-Language-Action)

• VLA Architecture Evolution — From RT-2 onward, cut along ablation axes rather than a timeline.
• Aerial VLA Models — Purpose-built architectures and training recipes (VLA-AN, CognitiveDrone, UAV-Track VLA, AutoFly).
• Language-Conditioned Flight Control — UAV-Flow, VLN-Pilot, and two 2026 aerial vision-language-navigation roadmaps.
• Foundation-Model-Assisted Planning — CoDrone, FM-Planner, NavFoM, FlyMirage: where the LLM sits in the loop.
• Onboard Deployment — Inference speedups, edge budgets, and action age as the real lower bound on control rate.
• Action Heads and Action Chunking — Discrete tokens, regression, diffusion, flow matching — and how to pick chunk length.
• Data, Pretraining and Cross-Embodiment — OXE/DROID, co-training, latent actions; flight data scarcity as the first constraint.
• RL Post-Training and Self-Improvement — GRPO/PPO on top of a VLA, where rewards come from, and how reward hacking shows up.
• Benchmarks and Reporting — LIBERO to RoboArena; the three choices hiding behind a single success-rate number.
• World-Model-Augmented VLA — A world model as data source, policy component, simulator, or evaluation proxy — four different claims.

4. VLMs

• Remote-Sensing VLMs — GeoChat, RSGPT, SkySenseGPT and why overhead imagery breaks ordinary VLMs.
• Aerial Scene Understanding Benchmarks — UAVBench, BEDI, and what "understanding" is actually scored as.
• LLM-Powered UAV Agents — CityNavAgent, ACDC: language models as planners and tool users for flight.
• Edge VLM Deployment — Lightweighting, distillation, BLIP-2, and what survives an 8 GB budget.
• General VLM Architecture and Vision Encoders — Four components, CLIP to native resolution, connectors, and the token budget.
• VLM Instruction Tuning and Alignment — Building instruction data, the limits of LoRA, frozen versus joint training.
• VLM Evaluation and Hallucination — MMMU, MMBench, POPE, plus negative sampling and yes-bias.

5. Survey Close-Read

A five-part close reading of a single survey (arXiv:2605.00080), section by section.

• Survey Overview and Structure
• World Model as Policy
• World Model as Simulator
• Video Generation World Models
• Benchmarks and Evaluation

6. Paper Cards

Short annotated cards — one per paper, with the claim, the setup, and the caveat.

• World-Model Paper Cards — 13 papers.
• VLA Paper Cards — 11 papers.
• VLM Paper Cards — 12 papers.
• Benchmark and Dataset Papers — 9 papers.

7. Hands-On Guides

• Environment Setup — CUDA, PyTorch, ROS2, and the version pairings that actually work.
• Reproducing UAV-Flow — Fine-grained language-conditioned drone control in the real world.
• Reproducing CognitiveDrone — Data collection and benchmark for cognitive UAV tasks.
• Reproducing FlightDiffusion — Diffusion-generated FPV video, plus two corrections to the paper.
• Reproducing MotionScape — A benchmark for future-video generation from a drone's point of view.
• Reproducing GeoChat — Remote-sensing VLM.
• Reproducing DreamerV3-Drone — World-model-based drone racing.
• Reproducible Project Candidates — 17 projects that one person with one machine can actually run.
• Reproducing DeepDrone — Natural-language drone control with an LLM.
• Reproducing RemoteCLIP — A remote-sensing vision-language foundation model.
• Reproducing Flightmare — A high-fidelity drone simulator.

8. Research Frontier

• Research Roadmap — A five-step path from learner to researcher, with the FINER criteria applied at each step.
• Open Problems — Where aerial VLA, VLM and world models are genuinely under-served.
• Critical Reading — The CRITIC method and a reading-note template.
• Latest Progress and Lab Tracking — 2025–2026 progress, and how to keep tracking it yourself.

9. Self-Assessment

Questions only — the body text gives pointers, never answers.

• VLA Self-Assessment — Three tiers of questions plus 12 aerial-specific ones.
• World-Model Self-Assessment
• VLM Self-Assessment

Mind Maps

• Field Overview — How world models, VLA and VLM relate.
• World-Model Taxonomy
• VLA Evolution Timeline
• Recommended Reading Order

References

• Full Paper List — 195 entries with a ★★★ / ★★☆ / ★☆☆ recommendation scale.
• VLA Watch, 2026-10 — Incremental frontier log.
• World-Model Watch, 2026-10
• VLM Watch, 2026-10
• Awesome-List Annotations — Notes on NTUMARS/Awesome-World-Model-for-Robotics-Policy and what we add on top.

---

How the numbers stay trustworthy

One rule runs through the whole repository, and it is enforced by tooling rather than by good intentions: every number comes from a real run. The demos under code/ are CPU-only and take about two minutes each; if a figure appears in the text, the script that produced it is in the repository.

Known limit, stated rather than hidden: the reference hardware is a single laptop GPU with 8 GB of VRAM, so nothing here trains a world model or a VLA from scratch. Anything that would need more is marked as such instead of being faked.

Citation

See CITATION.cff in the repository root — GitHub renders it as a "Cite this repository" button with BibTeX and APA exports.

Contributing

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