# Xbotics Embodied Guide: a Chinese-language roadmap through embodied AI, structured as eleven chapters > A documentation repository rather than a codebase, with ten 4 to 8 week learning tracks spanning RL, imitation learning, 3D vision, VLA and Sim2Real, plus simulation setup notes, a 120-person researcher index and a company map. Everything is written in Chinese. **Xbotics-Embodied-AI-club/Xbotics-Embodied-Guide** — Xbotics community guide to embodied AI: connects surveys, learning paths, simulation study, open-source hardware, interviews and a company landscape to help newcomers and practitioners quickly find a path, ship projects and join open source. - Repository: https://github.com/Xbotics-Embodied-AI-club/Xbotics-Embodied-Guide - Stars: 1,448 · Forks: 109 - Language: Python - License: NOASSERTION - Published: 2026-10-07 · Updated: 2026-10-07 · Language: en - Canonical page: https://hysenlabs.com/projects/xbotics-embodied-ai-club-xbotics-embodied-guide ## Documentation with a curriculum shape, not a link dump The repository's own description states its thesis plainly: it strings together survey, learning path, simulation study, open source hardware, people and companies, to help newcomers and practitioners locate a path, ship projects and join open source. It also says what it is not trying to be. It does not aim to be an encyclopedia; the goal is a clear path plus engineering practice. What that produces is a curriculum rather than a bookmark collection. The Learning Map section is a table of eleven numbered chapters, each with a key-content summary and a link into `docs/`. Chapter one is the survey, covering terminology, manipulation, world models, locomotion control and navigation as a perception-to-decision-to-control panorama. Chapter three is foundations, running from coordinate frames and kinematics through Transformers, diffusion and engineering environment setup. Chapter four is classical methods: behaviour cloning, DAgger, GAIL, value functions, policy gradients, Actor-Critic, SigLIP and CLIP, iLQR, MPPI and MPC. Chapter five is the current frontier, covering VLA models, Diffusion Policy, Sim2Real, supervised fine-tuning of large models, and 2023 to 2025 progress with evaluation. The Quick Start section is deliberately short, because there is nothing to install. The repository is documentation, so the entire setup is a clone: ```bash # 1. clone the repository git clone https://github.com/Xbotics-Embodied-AI-club/Xbotics-Embodied-Guide.git cd Xbotics-Embodied-Guide ``` After that, step two is opening `docs/1-embodied-overview/README.md` for the big picture and picking one track from `docs/2-roadmaps/README.md`, and step three is working the track and reporting back through an Issue. The ordering is the argument. Survey before foundations before classical methods before frontier means you meet a method after the vocabulary and the underlying mathematics, not before. A reader who arrives knowing PyTorch but not why Sim2Real is hard will find chapter five confusing; a reader who follows the order will not. Two directories sit alongside `docs/`. There is `files/`, which the career section points at as a place for organised drafts and source material, and a `news/` directory. There is also a `.cursor/` directory, which is an editor rules folder rather than anything about robotics, a useful signal about who maintains the repository. ## Ten tracks, each with an acceptance check The Learning Map describes chapter two as ten learning tracks, and the description of their structure is the most useful sentence in the README: prerequisites, then a 4 to 8 week plan, then milestones, then acceptance. That last element is what separates this from a topic list. The named tracks cover reinforcement learning, imitation learning, 3D vision, planning and control, localisation and navigation, tactile sensing, VLA, Sim2Real, world models, and the data flywheel. The Quick Start section tells you to pick one of these tracks and mentions the same four to eight week duration. Each section of the documentation also opens with a three-things-to-start motif, which the project philosophy section describes as not requiring you to begin with long tutorials. That is a real editorial choice. Most overviews open with a prerequisite reading list nobody finishes; opening with three concrete actions means you can be running something inside a session. The audience section names four groups, and they map onto the tracks rather than onto chapters: Python and control fundamentals learners, university students and graduate students who need a course project, thesis or paper reproduction route, algorithm and robotics engineers moving vision, RL or large models into embodied work, and hardware and open source community participants who care about the full chain from simulation to deployment to data. The acceptance criterion structure deserves one more word, because it is the mechanism that makes a roadmap falsifiable. A track that ends in an acceptance check can be failed, and a failed track tells you what you do not yet know. Very few learning paths in this field are built that way, and it is the single thing here worth borrowing for your own study plan. ## Simulation and hardware chapters name specific tools Chapter six covers simulation and the README lists the simulators by name: Isaac Lab, MuJoCo, PyBullet, Genesis and Gazebo. The scope is described as building virtual environments and training models, which is the right framing since the difficulty in that chapter is environment setup rather than algorithm selection. Chapter seven covers open source hardware and is anchored on LeRobot, the Hugging Face framework, with pretrained models, datasets, hardware support and Sim2Real deployment. That is a considered choice of anchor. LeRobot is the largest common denominator between academic and hobbyist manipulation work, and a guide that anchors there is giving you something you can actually buy a dataset for. Two other chapters are unusual enough to call out. Chapter eight is a people index, described as roughly 120 people across manipulation, humanoid and VLA work, split between academic and industry. Chapter nine is a company map organised by hardware whole units, key components, algorithm platforms and data services, with humanoid, quadruped and mobile-plus-manipulation categories. These two chapters will date faster than the rest of the repository. A list of 120 researchers is accurate when written and stale within a year, because the field's visibility shifts annually and hiring changes faster than methods. Treat them as a starting index rather than a reference, and check dates before you act on them. The same caution applies to the `news/` directory. Together, chapters six and seven are where a reader should be most sceptical of the guide's own framing. The Sim2Real gap between a policy that scores well in Isaac Lab and one that works on hardware is the hardest problem in embodied AI, and no roadmap document closes it. What this repository can honestly offer is the list of tools and the ordering in which to try them. ## A licence badge that disagrees with the licence field Here is a factual discrepancy inside this repository that affects anyone who wants to reuse the content. The README carries a Creative Commons badge reading CC BY 4.0, and a `LICENSE` file exists at the repository root. But GitHub's own licence detection returns no recognised value for this repository, so the licence field in the repository metadata is unset rather than set to Creative Commons Attribution 4.0. Both statements are present at the same time. Unlike a code licence, this matters more than usual for a documentation repository, because CC BY 4.0 is the permissive-but-attribution-required licence: you may copy, adapt and redistribute the material, including commercially, provided you give appropriate credit, link to the licence, and indicate whether you made changes. That is a real grant for a curriculum you might want to translate, fork for a cohort, or fold into an internal onboarding programme. The practical resolution is the same as reading the file: the badge plus the `LICENSE` file are explicit statements by the project, and the metadata field is an absence rather than a contradiction. But because documentation licences are the kind of thing automated tooling checks, a repository-scanning tool in your organisation will likely flag this as unlicensed and block it. Read `LICENSE`, and treat the content as CC BY 4.0 with attribution. One more detail on permissions. The Quick Start section asks readers to raise an Issue for questions and points to `docs/10-contributing/README.md` for contribution rules, with chapter ten and eleven covering AMA, pull request rules, directory conventions and licensing. The contribution guide is the authoritative place for what the project wants changed, and it is worth reading before proposing an edit to a chapter. ## The career chapter, and why its vocabulary list is the useful part One chapter sits outside the technical sequence: a job-seeking roadmap aimed at campus recruitment, internships and career changers. It runs in five stages: pick a direction in one week, build capability over four to eight weeks, polish a portfolio over two to four weeks, build industry awareness continuously, then apply and interview on a rolling basis. What makes it worth reading is not the job advice, it is the vocabulary list. The guide publishes keyword sets by direction, and these are the terms a job description in this field actually uses. For VLA and manipulation: OpenVLA, pi0, Diffusion Policy, LIBERO, Bridge, action tokenisation, LoRA fine-tuning. For locomotion control and planning: MPC, iLQR, trajectory optimisation, Sim2Real, Isaac Lab, MuJoCo, whole-body control. For navigation and vision-language navigation: VLN-CE, Habitat, R2R, semantic maps, hierarchical planning. For data and engineering: teleoperation data collection, LeRobot, the data flywheel, evaluation scripts, W&B or TensorBoard. That list is a legitimate service even if you are not job hunting. It is a compact, current index of what the field considers its core components, and reading it tells you which projects are mainstream and which are fringe. The stage table is also honest about where the work is. Stage two asks you to get a minimal demo running on one track; stage three asks you to reproduce one frontier baseline, whether VLA, Diffusion Policy or navigation, and to put metrics and a demo video on a GitHub README. Reproducing a published baseline is the field's actual portfolio currency, and the guide is telling you that a course project will not carry as much weight as a reproduced number. One dependency to note: the career chapter repeatedly links to a separate job-information repository, Xbotics-Embodied-AI-Job, for postings and internships. That repository is outside this one, so its freshness is not something the main repository controls. ## Conclusion Xbotics Embodied Guide is worth an afternoon from anyone entering embodied AI who finds the field's problem statement harder than any individual algorithm. The structure is the contribution: a terminology overview first, then foundations, then classical methods, then the current frontier, then simulators, then real hardware, and a set of ten tracks that each resolve into prerequisites, a 4 to 8 week plan, milestones and an acceptance check. Two limits to weigh. Everything is in Chinese, so an English-only reader needs a translator for the whole thing rather than a paragraph or two, and the repository contains no code, so the Sim2Real gap between finishing a track and running a policy on hardware is entirely yours. Clone it, read `docs/1-embodied-overview/README.md` for the vocabulary, then pick one track from `docs/2-roadmaps/README.md` and check its acceptance criteria before you start, because those criteria are the only thing in the repository that tells you whether you actually learned the thing. ## FAQ ### What is the Xbotics Embodied Guide and who is it for? It is a community learning guide for embodied AI that connects a survey, learning tracks, simulation, open source hardware, a researcher index and a company map. The stated audience is Python and control fundamentals learners, students needing a course project or paper reproduction, engineers moving vision or RL into embodied work, and hardware community participants. It is written entirely in Chinese. ### What learning tracks does the guide offer and how long are they? Ten tracks, each built as prerequisites, then a 4 to 8 week plan, then milestones, then an acceptance check. The named topics cover reinforcement learning, imitation learning, 3D vision, planning and control, localisation and navigation, tactile sensing, VLA, Sim2Real, world models, and the data flywheel. The acceptance step is what makes a track falsifiable rather than a reading list. ### What licence applies to the Xbotics Embodied Guide content? The README carries a CC BY 4.0 badge and a LICENSE file at the repository root, while GitHub's licence detection returns no recognised value, so the licence field is unset. Read LICENSE as authoritative: Creative Commons Attribution 4.0 permits copying, adaptation and commercial redistribution provided you give credit, link the licence and indicate changes. That matters because documentation licences are what automated compliance tooling checks. ## Sources - [Issues](https://github.com/Xbotics-Embodied-AI-club/Xbotics-Embodied-Guide/issues) - [README](https://github.com/Xbotics-Embodied-AI-club/Xbotics-Embodied-Guide/blob/main/README.md) - [Xbotics-Embodied-AI-club/Xbotics-Embodied-Guide on GitHub](https://github.com/Xbotics-Embodied-AI-club/Xbotics-Embodied-Guide) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/xbotics-embodied-ai-club-xbotics-embodied-guide