Embodied-AI-Guide: A Chinese Knowledge Base for Embodied AI, and the RoboTwin 2.0 Tutorial That Anchors It

What Embodied-AI-Guide Is, and What It Is Not

The README describes the project as a Chinese knowledge base and resource index for embodied AI, positioned as an encyclopedia. The repository layout backs that up: the top level holds a README, a files directory, a topics directory, a LICENSE and a .gitignore. There is no package manifest, no source tree, no build script. You cannot install this project because there is nothing to install. What you get is a structured document plus a set of linked PDFs, videos, papers and external repositories.

The audience is explicit. The authors call themselves a team of embodied AI beginners writing from their own learning experience to help later arrivals. The guide is split into a hands-on chapter and an algorithm chapter, with a stated design principle of being brief in the practical part and encyclopedic in the reference part. That combination is unusual. Most beginner resources pick one mode: a tutorial that walks you through a single pipeline, or a link dump with no ordering. This one tries to do both, and the seams show in places, but the ordering is the point.

The RoboTwin 2.0 Lifecycle Tutorial at the Center of the Guide

The practical chapter is built entirely around RoboTwin 2.0, a dual-arm simulation platform built on SAPIEN. The README states the tutorial should take about a week and that completing it needs a GPU with at least 16GB of VRAM. That is a hard constraint, not a suggestion, and it is the first thing to check before committing to the exercise.

The tutorial is organized as a policy lifecycle rather than a feature tour. The reader first reads the RoboTwin 2.0 paper and learns what information a synthetic robot data sample requires, then installs the platform and collects data for one task, then trains a policy, then evaluates it. The guide names three questions it wants the reader to be able to answer by the end: where data comes from, how a policy is designed, and how model performance is measured. It also lists the problems with each data source, including the cost of real-robot collection, the low information density of human video, the sim-to-real gap in synthetic data, and hallucination in world-model-generated data. Naming those trade-offs rather than presenting simulation as a free lunch is the most useful editorial choice in the document.

Installing RoboTwin 2.0 and Collecting a First Dataset

The guide does not host installation instructions. It links to the RoboTwin 2.0 documentation for environment setup, and the README's own instructions assume that setup is already done. What the README does give is the data collection command and its arguments.

bash collect_data.sh ${task_name} ${task_config} ${gpu_id}
## Clean Data Example: bash collect_data.sh beat_block_hammer demo_clean 0
## Radomized Data Example: bash collect_data.sh beat_block_hammer demo_randomized 0

The script takes a task name, a task configuration and a GPU index. The README's worked example collects 50 episodes of the beat_block_hammer task, in either demo_clean or demo_randomized form, on GPU 0. The two configurations are the interesting part: demo_clean is the fixed setup used for the later evaluation, while demo_randomized varies the scene. Collecting both is what lets you see how much of a policy's success comes from the task itself versus from the specific placement it was trained on. The README does not document what the script prints, how long collection takes, or where the data lands.

Training ACT and Reading the 56 Percent Baseline

After collection, the guide points to a separate RoboTwin 2.0 tutorial for reproducing ACT, described as a classic manipulation policy. The README states that training ACT takes roughly 12GB of VRAM, which is lower than the 16GB floor given for the tutorial as a whole, so a 16GB card has headroom for the evaluation step as well.

Evaluation happens under demo_clean, and the README reports an ACT success rate of about 56 percent, pointing to the RoboTwin 2.0 leaderboard for the full picture. That number is worth pausing on. It is a baseline, not a target, and it is exactly the kind of figure a beginner needs in order to tell whether their own reproduction has gone wrong or is simply matching the reference. The guide does not explain how success rate is computed or what the variance across seeds looks like, so treat 56 percent as a checkpoint to compare against rather than a specification.

The Reference Chapter: Link Indexes and Where They Thin Out

The third chapter collects orientation material: a technical roadmap PDF with a companion bilibili video, a Stanford robotics introduction, a systems-oriented blog, Chinese-language WeChat accounts, Xiaohongshu accounts, laboratory summaries on Zhihu, a Chinese highly-cited researcher list, and a long set of conference and journal names spanning robotics, vision, learning and language venues. A fourth chapter organizes algorithms from engineering and geometry tools up through 2D, 3D and 4D visual representation, then to reinforcement learning, imitation learning, vision-language-action models and planner-based systems.

The weakness is that this half is an index, not an explanation. Many entries are a name and a link with no annotation about difficulty, prerequisites or why a particular item sits where it does. The structure implies an ordering, but a reader who does not already know the field cannot tell which of two adjacent papers to read first. The guide is honest about being written by beginners, and this is where that shows most. It is a good map and a thin commentary.

Where Embodied-AI-Guide Is the Wrong Tool

If you want to run something today, this repository will not help you directly. It contains documentation and links, and the actual executable work lives in RoboTwin 2.0 and the other projects it points to. You will be juggling several external repositories, each with its own environment requirements, and the guide does not attempt to reconcile them into one setup.

Language is the second boundary. The README, the roadmap, the community list and the algorithm chapter are in Chinese. English readers can follow the linked papers and the RoboTwin documentation, but the connective tissue that explains why topics are ordered the way they are is not available in English. There is also a hardware boundary: the practical chapter assumes a discrete GPU with at least 16GB of VRAM, so anyone on a laptop or a CPU-only machine gets the reading list and nothing else. Finally, the guide targets manipulation in simulation. If your interest is navigation, legged locomotion or deployment on physical hardware, the tutorial path does not cover it.

How It Differs from a Curated Paper List

The closest comparison is a community-maintained awesome list, such as the paper lists the guide itself links to for humanoid robot learning or vision-language-action models. Those repositories are flat collections: a category heading, then entries. They are excellent for scanning a subfield and poor for sequencing, because nothing tells a newcomer where to begin or when to stop.

Embodied-AI-Guide inverts that. Its practical chapter picks one platform, one task family and one policy, and walks a full cycle from data collection to a reported success rate. That gives a beginner a finish line. The cost is coverage: a single pipeline cannot show you the breadth of the field, which is why the reference chapters exist alongside it. The two halves are doing different jobs, and reading the guide as if either half alone were the whole thing will mislead you. Use the RoboTwin chapter to build intuition and the reference chapters to find out what you have not seen yet.

Licence Status and What to Verify Before Adopting It

The repository metadata reports the licence as NOASSERTION, meaning no standard licence identifier was detected. The LICENSE file exists at the top level, so there is a licence document, but its terms are not summarized in the README and the repository does not present an SPDX identifier. If you plan to reuse the guide's text, redistribute the bundled PDF in the files directory, or build teaching material from it, read that LICENSE file yourself. The guide is a documentation project, so the licence question is about reuse of prose and bundled assets rather than about linking code into your application. Nothing here constitutes legal advice.

On maintenance, the last push to the repository was on 2026-09-08, and the repository is not archived. The most recent release is tagged v1.0.0, dated 2026-02-20. The README's news section records a reorganization completed on 2026-01-15, which means the current structure is recent and older links or bookmarks may point at sections that moved. The upgrade cost is therefore low in the software sense, since there is nothing to upgrade, but non-zero in the reading sense: the table of contents has changed once already, and the guide's value depends on the external links it points to staying alive. Those links are outside the project's control.

Editorial conclusion

Adopt Embodied-AI-Guide if you are a Chinese-reading newcomer who wants a structured entry point and a concrete RoboTwin 2.0 exercise, or an instructor assembling a reading list. Do not adopt it if you need runnable code, an English-language resource, or a maintained software dependency, since the repository contains only a README, a files directory, a topics directory, a LICENSE and a .gitignore. Before relying on it, open the RoboTwin 2.0 install tutorial it links to and confirm you have a GPU with at least 16GB of VRAM, because the guide states that is the floor for the tutorial.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. Releases
5. TianxingChen/Embodied-AI-Guide on GitHub