Awesome MuJoCo

A collection of awesome projects using MuJoCo.

Simulation

• Official MuJoCo: Official MuJoCo repository.
• Genesis: Taichi-based MuJoCo implementation with faster contact detection.
• Mujoco Warp: Nvidia Warp-based MuJoCo implementation. (under construction)

Control

• MuJoCo-MPC: Official MuJoCo MPC library by Google DeepMind. Provides efficient real-time model predictive control with parallel GPU implementation.
• Mink: Fast and robust inverse kinematics solver using MuJoCo's physics engine. Supports various robot configurations.
• Mjinx: Modern inverse kinematics library built on MJX (JAX-based MuJoCo). Enables gradient-based optimization.
• mjctrl: Comprehensive robot manipulator control library with various control schemes (position, force, impedance control).
• Generative Predictive Control: Novel MPC approach using flow matching for policy generation. Built on MJX for efficient parallel execution.
• Model-based-Diffusion: Trajectory optimization using diffusion models and sampling. Leverages MJX for parallel simulation.
• DIAL-MPC: High-performance MPC framework for legged robots. Features GPU acceleration and sim2real transfer capabilities.
• Whole-Body-MPPI: Real-time whole-body control for legged robots using Model Predictive Path Integral control.
• Hydrax: Efficient sampling-based MPC implementation using MJX. Suitable for real-time control applications.
• DROP: Advanced in-hand manipulation framework combining MuJoCo-MPC with sampling-based optimization for robust object manipulation.

Learning

• MuJoCo-Playground: robot learning library and sim2real interface by Google DeepMind.
• Brax: parallel simulation support MJX with common RL algorithms.
• DM-Control: CPU-based RL environment library by DeepMind.
• LocoMuJoCo: imitation learning library for legged robots.
• so101-nexus: Full-stack robot learning for the SO-101 arm (teleoperation, imitation learning, RL) in MuJoCo, with an optional NVIDIA Warp backend.

Teleoperation

• MujocoAR: use iOS device to control MuJoCo frames in real-time.

Tooling

• mjcf-doctor: Linter for the model errors MuJoCo accepts silently — collision masks that let everything fall through the floor, contacts soft enough to swallow a peg, inertia off by 10^6 from a kg·mm² export. Also has a batch audit mode; running it over all 82 Menagerie models found a 48 mm self-collision in hello_robot_stretch_3 that fires 398 kN at t=0.

Assets

• MuJoCo-Menagerie: collection of MuJoCo models.
• Unitree-MuJoCo: Official Unitree MuJoCo models.