Autopilot-Notes: a Chinese-language autonomous driving knowledge base in Markdown

What Autopilot-Notes actually is, and who it is written for

The README describes the repository as a systematic open knowledge base for learning autonomous driving technology, covering the full stack from perception and localisation to planning and control. Every top-level entry is a chapter directory: ch01_基础 for fundamentals, ch02_硬件 for hardware, ch03_感知 for perception, ch04_定位 for localisation, ch05_策略规划 for prediction and planning, ch06_控制 for control, ch07_产品 for product topics, ch08_工具 for tooling, ch09_厂商方案 for vendor solutions, and ch10_每日前沿 for the daily digest. The primary language listed for the repository is Shell, which reflects the scripts/ directory and the automation around the digest rather than the content itself.

The intended reader is a developer moving from general software or machine learning into autonomous driving, or an engineer who already works in one module and wants the vocabulary of the neighbouring ones. The README frames the goal as helping developers go from beginner to advanced. That framing matters: a systems engineer who needs a production planner will not find one here, and the repository does not claim otherwise. What it offers is the map, not the vehicle.

The chapter map: from coordinate systems to VLA

The structure is the product. ch01 opens with coordinate systems, camera intrinsics and extrinsics, Kalman filtering, image transforms, 3D reconstruction including NeRF, datasets, Transformer, NLP, neural architecture search and reinforcement learning. That ordering is deliberate: the notes assume you will meet rotation matrices and probability before you meet a detector.

ch03 is the densest algorithmic chapter: 2D detection, 3D detection, BEV, Occupancy Network, and a section on end-to-end driving and VLA. The README's own architecture table maps each layer to the question it answers, with hardware as seeing and computing, perception as what was seen, localisation as where am I, prediction as where it is going, planning as how should I move, and control as how to execute. The control chapter is narrow on purpose: PID, LQR and MPC, one file each.

The README also draws the paradigm shift from modular pipelines to end-to-end and VLA models, contrasting a decoupled stack with a single large model and a vision-language-action arrangement where a language model sits between the visual encoder and an action head. It names Tesla FSD V13, Huawei ADS 4.0, Li Auto MindVLA and Waymo EMMA as examples of those paradigms. Those are attributions in the README, not benchmarks, and the repository does not publish measurements behind them.

How to start reading it: clone, then follow the path

There is no package to install. The README's quick start is a learning path diagram, not a setup script, so the first real action is cloning the repository and opening the chapter you need.

git clone https://github.com/gotonote/Autopilot-Notes.git
cd Autopilot-Notes
ls

The listing should show the chapter directories ch01_基础 through ch10_每日前沿 alongside LICENSE, README.md, imgs/ and scripts/. From there, the README's recommended order is 阶段1 基础, then 硬件, 感知, 定位, 规划, 控制, 产品, 工具. If you want the newest material first, go to the digest instead.

ls ch10_每日前沿

You should see dated Markdown files such as 2026-09-10.md, 2026-09-09.md, 2026-09-08.md and 2026-09-03.md, plus an archive directory. The README states that the digest updates automatically at 19:00 on weekdays and that files older than seven days are moved into the archive. Note the README's own caveat: the home index is generated between marker comments, so if you edit it by hand the next push will overwrite you.

For the one chapter with runnable instructions, ch08 points at a CARLA section covering installation, the Python API, sensor configuration, scenario construction and worked examples. That is the closest thing to a hands-on track in the repository, and it is still notes about CARLA rather than a CARLA distribution.

The daily digest is the part with an upkeep bill

ch10_每日前沿 is the only component with a stated schedule. The README says it publishes at 19:00 on weekdays, that entries are tagged with keywords such as VLA, World Model, Occupancy Network, Robotaxi and 无图智驾, and that anything older than seven days is archived. The most recent push to the repository was on 2026-09-10, and the newest digest file listed in the README index is 2026-09-10.md. The README describes the schedule; whether the automation has kept to it since is something you check by looking at the commit history, not something the README can tell you.

The keyword tagging is genuinely useful for filtering: if you only care about BEV or occupancy networks, you can scan the index table rather than open every file. The cost is that the digest is a moving target. A cloned copy ages the moment the automation stops, and the archive directory means the index you see is a seven-day window rather than the full history. If you want the digest as a durable record, mirror the archive yourself.

Where it is the wrong tool

The repository does not ship code you can import. There is no Python package, no model weights, no training script and no inference server in the top-level listing. If your task is to fine-tune a BEV detector, evaluate a planner on nuScenes, or deploy a network with TensorRT, Autopilot-Notes gives you the concepts and the vendor context, and you will still need the upstream project for every one of those steps. The topics list includes tensorrt and camera-calibration, but a topic tag on a notes repository is not a library.

The second limitation is language. The README, the chapter names and the file names are in Chinese, with English terms mixed in for technical vocabulary. An English-only reader can follow the file paths and the code-adjacent sections but will lose most of the prose, and the repository does not document a translated edition.

The third is verification. The vendor comparisons, the paradigm tables and the digest entries are summaries. The README attributes end-to-end and VLA approaches to specific companies, but it publishes no test methodology, no dataset and no reproducible numbers. Treat every performance-flavoured claim in the notes as a pointer to the original source, not as an evaluated result.

Alternatives, and what changes if you pick one

The obvious alternative is the upstream documentation of the tools the notes describe. CARLA publishes its own install guide and Python API reference, and the ch08 notes are a summary of that material. Choosing CARLA's docs over Autopilot-Notes gives you version-accurate instructions and an issue tracker; choosing Autopilot-Notes gives you the surrounding context, such as why a sensor configuration matters for the perception chapter you read earlier.

A second alternative is a university course or a textbook on autonomous driving. Those give a graded sequence with exercises and a consistent notation. Autopilot-Notes gives breadth and currency instead: it will mention a 2026 vendor announcement that no textbook has, and it will mention it in a table row rather than a chapter. The trade is depth for freshness, and the repository's own structure admits this by keeping the digest in a separate chapter from the fundamentals.

A third option is a vendor's own engineering blog. Those are written by the teams that shipped the system and carry details the notes can only summarise. They are also narrower and, in most cases, promotional. If you are deciding between reading one vendor blog and reading ch09_厂商方案, the honest answer is that the notes will help you ask better questions of the blog.

Licence, contribution and what to check before you depend on it

The repository is licensed Apache-2.0, and the LICENSE file sits at the top level. That permits reuse and redistribution under the licence's conditions, including attribution and the patent grant; it does not settle the rights in third-party figures, vendor slides or quoted material that the notes may reference. If you plan to republish chapter content, check the provenance of images under imgs/ and of any quoted vendor material yourself. Nothing here is legal advice.

The README states that pull requests are welcome, and the repository contains a 文章撰写规范.md file at the top level, which indicates a house style for contributed articles. If you intend to submit, read that file before writing; a contribution that ignores it is likely to be reworked.

Before depending on Autopilot-Notes for anything time-sensitive, verify three things in the clone rather than in the README: the date of the newest file in ch10_每日前沿, the commit date of the chapter you intend to cite, and whether the ch08 CARLA instructions match your installed CARLA version. Those three checks take a few minutes and they are the difference between citing a maintained note and citing a snapshot.

Editorial conclusion

Adopt Autopilot-Notes if you want a Chinese-language reading path through the ADAS stack, from coordinate systems and Kalman filters up to BEV, occupancy networks and VLA, and you are willing to follow the chapter links by hand. Do not adopt it if you need a library, a simulator or an SDK: the repository is Markdown plus a scripts directory, and the only executable surface the README points at is the CARLA material under ch08. Before relying on it, clone it and check two things yourself: whether the ch10 daily digest files are still being pushed, and whether the ch08 CARLA notes match the CARLA release you actually have installed.

Frequently asked questions

Official sources

1. gotonote/Autopilot-Notes on GitHub
2. Issues
3. License: Apache-2.0
4. README