Embodied AI Starter Map | 具身智能入门.png)
What is this?
A Chinese-language 'Embodied AI Starter Map' that organizes a learning path from robotics fundamentals (coordinates, kinematics, ROS 2) up to VLA policies, World Models, model-based RL and World Action Models (WAM), linking papers, code, benchmarks and practice guides at each step.
A Chinese-language 'Embodied AI Starter Map' that organizes a learning path from robotics fundamentals (coordinates, kinematics, ROS 2) up to VLA policies, World Models, model-based RL and World Action Models (WAM), linking papers, code, benchmarks and practice guides at each step.Is this a runnable codebase or just a reading index?It is purely a documentation/index repository of markdown pages and links — the author says you can treat it as a reading index or follow tutorials starting with MuJoCo, Isaac Sim or ROS 2 Humble, but the actual runnable code lives in the external linked projects.
It is purely a documentation/index repository of markdown pages and links — the author says you can treat it as a reading index or follow tutorials starting with MuJoCo, Isaac Sim or ROS 2 Humble, but the actual runnable code lives in the external linked projects.Where should a beginner start?Start from the main route table and pick a track based on your goal — Robotics, VLA, World Model, WAM or RL/MBRL — each one links to an entry page like Robotics Basics or RL Basics; if unsure, start from the Knowledge Graph overview page.
Start from the main route table and pick a track based on your goal — Robotics, VLA, World Model, WAM or RL/MBRL — each one links to an entry page like Robotics Basics or RL Basics; if unsure, start from the Knowledge Graph overview page.Do I need ROS 2 or a simulator installed to follow this?The robotics chapters assume Ubuntu 22.04 with ROS 2 Humble as the baseline, and the hands-on tutorials use MuJoCo+Gymnasium or Isaac Sim; ROS 2/OpenCV install guidance is pointed to an external community forum rather than included here.
The robotics chapters assume Ubuntu 22.04 with ROS 2 Humble as the baseline, and the hands-on tutorials use MuJoCo+Gymnasium or Isaac Sim; ROS 2/OpenCV install guidance is pointed to an external community forum rather than included here.Does it include the actual model implementations, or just references?Just references — model and policy code (OpenVLA, OpenPI, Diffusion Policy, LeRobot, V-JEPA 2, TD-MPC2, DreamerV3, etc.) and benchmarks live in their original repos; for real hardware demos like the Piper arm or bimanual-vla, exact commands depend on the original project's own README.
Just references — model and policy code (OpenVLA, OpenPI, Diffusion Policy, LeRobot, V-JEPA 2, TD-MPC2, DreamerV3, etc.) and benchmarks live in their original repos; for real hardware demos like the Piper arm or bimanual-vla, exact commands depend on the original project's own README.Who is this guide meant for?Researchers and learners who already know some robotics basics (or want to pick them up) and want a structured path into VLA, World Models, MBRL and WAM, with papers, code and benchmarks mapped out for each direction.
Researchers and learners who already know some robotics basics (or want to pick them up) and want a structured path into VLA, World Models, MBRL and WAM, with papers, code and benchmarks mapped out for each direction.What is the difference between a World Model (WM) and a World Action Model (WAM)?VLA learns directly from vision and language to action; WM predicts future states conditioned on actions; WAM structurally feeds that future prediction back into action generation, tying future representation, training-time supervision and asynchronous execution together.
VLA learns directly from vision and language to action; WM predicts future states conditioned on actions; WAM structurally feeds that future prediction back into action generation, tying future representation, training-time supervision and asynchronous execution together.Is a World Model the same thing as video generation?No — the README explicitly states WM is not equal to video generation and MBRL does not require a video-based WM; whether something counts as a full robotic World Model depends on action-conditioning, future goals, and closed-loop evidence, not on generating video.
No — the README explicitly states WM is not equal to video generation and MBRL does not require a video-based WM; whether something counts as a full robotic World Model depends on action-conditioning, future goals, and closed-loop evidence, not on generating video.What benchmarks does the guide cover?It surveys LIBERO, CALVIN, ManiSkill, Meta-World, RoboDojo, RoboTwin, RoboCasa and DMControl, explaining each one's intended use and the boundaries for fair comparison between them.
It surveys LIBERO, CALVIN, ManiSkill, Meta-World, RoboDojo, RoboTwin, RoboCasa and DMControl, explaining each one's intended use and the boundaries for fair comparison between them.Can I follow a path from simulation to a real robot arm?Yes — there's a dedicated sim-to-real track covering eye-hand calibration, ros2control and rosbag2 leading to a real Piper arm under ROS 2 Humble, plus a bimanual-vla reference for dual-arm VLA deployment.
Yes — there's a dedicated sim-to-real track covering eye-hand calibration, ros2control and rosbag2 leading to a real Piper arm under ROS 2 Humble, plus a bimanual-vla reference for dual-arm VLA deployment.What are similar projects?Zero-WAM, transferlearning, Embodied-AI-Guide, WholebodyVLA, awesome-rsi, Code-as-World, video-prediction-policy-2, FluxVLA
Zero-WAM, transferlearning, Embodied-AI-Guide, WholebodyVLA, awesome-rsi, Code-as-World, video-prediction-policy-2, FluxVLAWhich topics is it in?vision-language-models, ros, world-action-model, world-models, roadmap, paper, robotics, simulation, reinforcement-learning, academic-research, learning-resources, benchmark
vision-language-models, ros, world-action-model, world-models, roadmap, paper, robotics, simulation, reinforcement-learning, academic-research, learning-resources, benchmark