MuJoCo Atlas#

Learning home · MuJoCo contents

Course text is primarily Chinese, with English terminology and summaries. The original content is preserved.

Understand MuJoCo through native APIs, physics concepts and versioned source code.

中文 · Series home · Introductory guide · Curriculum · Source map · Versions · Roadmap · Project tracker

Part of Sim Atlas, an independent community learning series with two complete planned tracks: applications (modeling, control, robotics, sensing and data) and principles/source (dynamics, contact, solvers, integration and extensions).

The introductory guide, pinned source map and E1 chapters are available in Chinese: modeling and frames and state and time. E1 covers A1/A2 and the foundations of B0/B4, including coordinate/inertia conventions, state ownership, snapshots, sampling phases and integration semantics. E2 now covers actuation and control, robotics and kinematics, and task interfaces, including multi-input actuator blocks, force clamping, bounded local IK and observation-driven task transitions. The full course is in development.

Original examples have only passed static Python/XML checks; they were not compiled or executed by MuJoCo. Validation and source discrepancies separate source review from runtime evidence. This phase prioritizes understanding engine subsystems. Minimal snippets support explanation and are explicitly marked when unexecuted. No new simulation campaigns, benchmarks, training or scoring are included; later experimental material will reuse DexLab with its original version and workload boundaries.

E2 examples received syntax/source review only: no model compilation, IK execution or grasping run was performed. E2 validation records the exact scope and next-stage dependencies.

E3 now covers dynamics and the step pipeline, contact models and material mixing, solvers and integration, and force observations. This completes the A4/B0–B5 core lessons, with version-specific convergence, warmstarts, discrete metrics, contact frames and sampling contracts. Original examples and E3 validation are source/static only; no native compilation, forward/step, collision or numerical experiment was run. Detailed flex/IPC internals and backend extensions are covered in E6.

E4 now covers sensors and sampling, cameras and geometry queries, and rendering and the viewer. A6 explains native sensor blocks, history, camera/ray conventions, classic RGB/depth/segmentation, context ownership, headless paths and passive synchronization. It records version-specific source discrepancies and the distinct Filament API. Original example and E4 validation are source/static only: no native import, model compilation, forward/step, graphics context, images or GUI were run. Physics backend internals are covered in E6; Filament coverage remains at the native-interface scope documented in E4.

E5 now covers CPU batching, MJX and device data, learning contracts and randomization, and recording and replay. A8/A9 and the batching/JIT/transfer/data part of B6 explain worker ownership, rollout initialization and warnings, native/MJX backend identities, task time and terminal observations, physical parameter dependencies, PRNG streams and checkpoint limits. JAX 0.7.2 is separately pinned for API source reading; it is not a tested MJX dependency lock. The original CPU example, 32 answered exercises and E5 validation are source/static only: no native import, compilation, rollout, device execution, training or data generation. Extension/backend internals continue in E6.

E6 covers callbacks and plugins, flex and elasticity, IPC contact, backend internals and differentiation, and multiphysics/source tracing. B6/B7 now connect state ownership, material forces, AL/CCD, partial advancement, actual AD paths and fluid approximations, with 40 answered reading exercises. Original flex snippet and E6 validation are source/static only: no native imports, model compilation, forward/step, JIT, rendering, gradients or experiments. The complete A0 installation course and whole-course review/DexLab evidence index remain E7.

Pinned upstream source · Attribution

Course source · GitHub · License