A robotics curriculum that actually runs on your laptop

What it does Every-Embodied is a structured, three-stage learning library for embodied AI. It starts with MuJoCo simulation and ROS basics, moves through computer vision and motion planning, and ends at reproducing frontier models like OpenVLA, SmolVLA, and Pi0. The target audience is deliberately broad: Python beginners, grad students, algorithm engineers looking to pivot into robotics, and educators who need a ready-made syllabus.

The interesting bit Most robotics courses drown you in kinematics derivations before you touch code. This one flips the pedagogy: run the demo first, then unpack the theory. The repo also chases the research frontier unusually aggressively for an educational project—its “SOTA reproduction” section includes VLN navigation, world models (RISE, LeWorldModel), soft-body simulation in SIM1, and even a Spring Festival dance-robot gag built from Doubao-generated videos.

Key highlights

• Hardware-agnostic progression: pure simulation (MuJoCo, Isaac Sim) up through real-arm teleoperation with LeRobot on a D-Robotics RDK-X5 board
• VLA/VLN coverage: OpenVLA, SmolVLA, Pi0, ETPNav, and drone LLM navigation with WebUI frontends
• 3D reconstruction and design pipelines: LingBot-Map for video-to-point-cloud, ForgeCAD for parametric CAD export (STEP/STL/3MF)
• Low-barrier RL: state-based training that reportedly runs on 6 GB GPUs via UniLab + MotrixSim
• CC-BY 4.0 license, actively maintained with dated news updates through mid-2026

Caveats

• The README is overwhelmingly in Chinese; English learners may need browser translation
• Scope is encyclopedic—21+ topical folders—so self-discipline is required to avoid yak-shaving
• Some “reproductions” are tutorial walkthroughs rather than drop-in reference implementations

Verdict Worth bookmarking if you’re a developer trying to bootstrap into embodied AI without a robotics PhD. Skip it if you need a single, production-hardened framework rather than a broad survey course.