16.5K ·GitHub
What is this?
A free, encyclopedia-style Chinese knowledge base and resource index for embodied AI — robots that perceive and act in the physical world — maintained by the Lumina Embodied AI community. It is one of the most popular embodied-AI guides in China, with over 16,000 GitHub stars.
A free, encyclopedia-style Chinese knowledge base and resource index for embodied AI — robots that perceive and act in the physical world — maintained by the Lumina Embodied AI community. It is one of the most popular embodied-AI guides in China, with over 16,000 GitHub stars.Is this an app I install, or something I read?Something you read. There is no software to run: it is a GitHub repo of structured markdown chapters plus curated links to courses, papers, simulators, benchmarks, datasets and labs; the only hands-on part relies on external tools like the RoboTwin 2.0 simulator.
Something you read. There is no software to run: it is a GitHub repo of structured markdown chapters plus curated links to courses, papers, simulators, benchmarks, datasets and labs; the only hands-on part relies on external tools like the RoboTwin 2.0 simulator.How am I meant to work through it?Start at the 'Start From Here' section: the core path is one hands-on project — about a week on the RoboTwin 2.0 simulation platform — paired with encyclopedia-style chapters to consult as reference while you go deeper.
Start at the 'Start From Here' section: the core path is one hands-on project — about a week on the RoboTwin 2.0 simulation platform — paired with encyclopedia-style chapters to consult as reference while you go deeper.Can I use the guide if I don't read Chinese?Not really — the guide itself is written in Chinese, as an explicitly Chinese-language knowledge base. Many of the resources it links out to, such as Stanford's robotics course and the research papers, are in English.
Not really — the guide itself is written in Chinese, as an explicitly Chinese-language knowledge base. Many of the resources it links out to, such as Stanford's robotics course and the research papers, are in English.Is this aimed at robotics researchers, or can a newcomer follow it?It targets newcomers — the maintainers are embodied-AI beginners themselves and want to help people build a mental map of the field quickly. The hands-on week does assume you can set up a sim environment and train a model.
It targets newcomers — the maintainers are embodied-AI beginners themselves and want to help people build a mental map of the field quickly. The hands-on week does assume you can set up a sim environment and train a model.Do I need a real robot or any special hardware to follow the tutorial?No real robot — the tutorial runs entirely in the RoboTwin 2.0 simulator. You do need a GPU with at least 16GB of VRAM for the full walkthrough; the ACT policy training step alone takes about 12GB.
No real robot — the tutorial runs entirely in the RoboTwin 2.0 simulator. You do need a GPU with at least 16GB of VRAM for the full walkthrough; the ACT policy training step alone takes about 12GB.What is RoboTwin 2.0 and what does the one-week tutorial have you do?RoboTwin 2.0 is a simulation platform for robot manipulation built around ALOHA-style hardware. In the tutorial you configure it, collect 50 demos of the beatblockhammer task, train the classic ACT imitation-learning policy (about a day, ~12GB of VRAM), and evaluate it — a reproduced ACT reaches roughly 56% success on the democlean benchmark.
RoboTwin 2.0 is a simulation platform for robot manipulation built around ALOHA-style hardware. In the tutorial you configure it, collect 50 demos of the beatblockhammer task, train the classic ACT imitation-learning policy (about a day, ~12GB of VRAM), and evaluate it — a reproduced ACT reaches roughly 56% success on the democlean benchmark.What do the reference chapters cover beyond the tutorial?Algorithms (vision foundation models, robot learning, VLA models, LLM+robotics, navigation), Infrastructure (simulators, benchmarks, datasets), Control (control theory, kinematics, SLAM, ROS) and Hardware (embedded systems, mechanical design, depth and tactile sensors, data-collection rigs), plus a Useful Info section with courses, labs, conferences and paper lists.
Algorithms (vision foundation models, robot learning, VLA models, LLM+robotics, navigation), Infrastructure (simulators, benchmarks, datasets), Control (control theory, kinematics, SLAM, ROS) and Hardware (embedded systems, mechanical design, depth and tactile sensors, data-collection rigs), plus a Useful Info section with courses, labs, conferences and paper lists.Is the guide still being updated?Yes — it was open-sourced in March 2025, passed 10,000 GitHub stars in December 2025, and a full reorganization of the guide was completed in January 2026.
Yes — it was open-sourced in March 2025, passed 10,000 GitHub stars in December 2025, and a full reorganization of the guide was completed in January 2026.Why did the authors create this guide?They are a team of embodied-AI beginners who wanted to pass on their own learning experience, helping newcomers build domain awareness quickly and speeding up adoption of the field. They welcome new contributors via PRs and the Lumina community.
They are a team of embodied-AI beginners who wanted to pass on their own learning experience, helping newcomers build domain awareness quickly and speeding up adoption of the field. They welcome new contributors via PRs and the Lumina community.What are similar projects?RoboICL, lerobot, Motus2, Zero-WAM, Embodied-AI-Start, WholebodyVLA, flex-pi, Zetta-Embodiment
RoboICL, lerobot, Motus2, Zero-WAM, Embodied-AI-Start, WholebodyVLA, flex-pi, Zetta-EmbodimentWhich topics is it in?robot, world-models, embodied-ai, vlm, robotics, simulation, embedded, reinforcement-learning, hardware, knowledge-base, guide, computer-vision
robot, world-models, embodied-ai, vlm, robotics, simulation, embedded, reinforcement-learning, hardware, knowledge-base, guide, computer-vision