The first MuJoCo physics engine running on Apple MLX. Numerically identical to C MuJoCo. Cartpole moves, pendulum swings, forces work. Ported from JAX to Apple MLX by RobotFlow Labs.
MJX is Google DeepMind's JAX-based GPU physics engine for MuJoCo. MJX-MLX replaces JAX with Apple MLX, enabling GPU-accelerated physics simulation on M1/M2/M3/M4/M5 Macs via Metal.
=== TEST 1: Cartpole with force ===
qpos=[0.489, -1.137] qvel=[9.34, -20.51]
Cart moved: PASS
=== TEST 2: Pendulum swing ===
angles: [0.004, 0.161, 0.304, 0.420, 0.498]
Pendulum swinging: PASS
=== TEST 3: C MuJoCo comparison (200-step pendulum) ===
C MuJoCo: 0.533293
MLX: 0.533293
Difference: 0.000000 <-- PERFECT MATCH
=== TEST 4: Performance ===
500 steps in 18.75s (26.7 steps/sec)
| Module | Files | Lines | Status |
|---|---|---|---|
| Foundation (types, math, dataclasses, io) | 4 | 2,831 | Working |
| Core dynamics (support, smooth, forward, passive, inverse) | 5 | ~3,500 | Working |
| Constraint system (constraint, solver) | 2 | ~2,500 | Working |
| Collision (primitive, convex, driver, sdf, bvh, types) | 6 | ~3,000 | Working |
| Sensors and rendering (sensor, ray, mesh, scan, derivative, render) | 6 | ~2,300 | Working |
| Total | 24 files | ~12,000 lines | Full physics step works |
- Model loading from MuJoCo XML into MLX arrays
- Full forward dynamics step: kinematics, mass matrix, constraints, solver, integration
- Actuator forces (motor control)
- Euler integration with quaternion support
- Contact constraint assembly and solving (CG/Newton with Cholesky)
- Passive forces (gravity, spring/damper skip for zero stiffness)
- Numerical output matches C MuJoCo to 6 decimal places
Porting 19,000 lines of JAX to MLX required fixing ~200 individual issues across the codebase:
| Category | Count | Example |
|---|---|---|
| numpy int indexing into MLX arrays | ~130 | m.jnt_bodyid[i] needs int() wrapper |
| numpy/MLX matmul type mismatch | ~20 | efc_J @ d.qvel where qvel is numpy |
_impl fallback pattern |
~50 | (m._impl or m).field for flattened fields |
| Backend check removal | ~15 | isinstance(m._impl, ModelMLX) guards |
| Missing MLX ops | 3 | .take() -> mx.take(), linalg -> CPU stream |
| Derived field computation | 4 | dof_hasfrictionloss -> inline frictionloss > 0 |
| Shape/padding guards | 5 | solref 0-dim -> 1-dim padding in _kbi |
The key insight: JAX and MLX have nearly identical array APIs, but MLX is stricter about types.
Every numpy.int64 used as an array index must be cast to Python int. Every numpy.ndarray
passed to an MLX operation must be wrapped in mx.array(). The translation is mechanical but
pervasive.
git clone https://github.com/RobotFlow-Labs/Mujoco-mlx.git
cd Mujoco-mlx
pip install mujoco mlx scipy
# Run physics simulation
PYTHONPATH=mjx python -c "
import sys; sys.path.insert(0, 'mjx')
import mujoco
from mujoco.mjx_mlx._src import io, forward
import mlx.core as mx, numpy as np
xml = '<mujoco><worldbody><body pos=\"0 0 2\"><joint type=\"hinge\" axis=\"0 1 0\"/><geom type=\"capsule\" fromto=\"0 0 0 0 0 -1\" size=\"0.04\" mass=\"1\"/></body></worldbody></mujoco>'
m = mujoco.MjModel.from_xml_string(xml)
m_mlx = io.put_model(m)
d_mlx = io.make_data(m)
d_mlx = d_mlx.replace(qvel=mx.array([2.0]))
for i in range(100):
d_mlx = forward.step(m_mlx, d_mlx)
mx.eval(d_mlx.qpos)
print(f'Pendulum angle after 100 steps: {float(np.array(d_mlx.qpos)[0]):.6f}')
"
# Run test suite
PYTHONPATH=mjx python mjx/mujoco/mjx_mlx/test_mlx_physics.pyThe MLX port lives in mjx/mujoco/mjx_mlx/ alongside the original mjx/mujoco/mjx/ (JAX).
The C MuJoCo engine is unchanged -- MJX-MLX only replaces the Python GPU physics layer.
| Repository | What |
|---|---|
| IsaacLab-mlx | RL framework on Apple Silicon (10 task families) |
| zed-sdk-mlx | ZED stereo camera SDK for macOS |
| zed-ros2-wrapper-mlx | ROS2 ZED wrapper for macOS |
MuJoCo stands for Multi-Joint dynamics with Contact. It is a general purpose physics engine that aims to facilitate research and development in robotics, biomechanics, graphics and animation, machine learning, and other areas which demand fast and accurate simulation of articulated structures interacting with their environment.
This repository is maintained by Google DeepMind.
MuJoCo has a C API and is intended for researchers and developers. The runtime simulation module is tuned to maximize performance and operates on low-level data structures that are preallocated by the built-in XML compiler. The library includes interactive visualization with a native GUI, rendered in OpenGL. MuJoCo further exposes a large number of utility functions for computing physics-related quantities.
We also provide Python bindings and a plug-in for the Unity game engine.
MuJoCo's documentation can be found at mujoco.readthedocs.io. Upcoming features due for the next release can be found in the changelog in the "latest" branch.
There are two easy ways to get started with MuJoCo:
-
Run
simulateon your machine. This video shows a screen capture ofsimulate, MuJoCo's native interactive viewer. Follow the steps described in the Getting Started section of the documentation to getsimulaterunning on your machine. -
Explore our online IPython notebooks. If you are a Python user, you might want to start with our tutorial notebooks running on Google Colab:
- The introductory tutorial teaches MuJoCo basics:
- The Model Editing tutorial shows how to create and edit models procedurally:
- The rollout tutorial shows how to use the multithreaded
rolloutmodule: - The LQR tutorial synthesizes a linear-quadratic controller, balancing a
humanoid on one leg:
- The least-squares tutorial explains how to use the Python-based nonlinear
least-squares solver:
- The MJX tutorial provides usage examples of
MuJoCo XLA, a branch of MuJoCo written in JAX:
- The differentiable physics tutorial trains locomotion policies with
analytical gradients automatically derived from MuJoCo's physics step:
Versioned releases are available as precompiled binaries from the GitHub releases page, built for Linux (x86-64 and AArch64), Windows (x86-64 only), and macOS (universal). This is the recommended way to use the software.
Users who wish to build MuJoCo from source should consult the build from
source section of the documentation. However, note that the commit at
the tip of the main branch may be unstable.
The native Python bindings, which come pre-packaged with a copy of MuJoCo, can be installed from PyPI via:
pip install mujocoNote that Pre-built Linux wheels target manylinux2014, see
here for compatible distributions. For more
information such as building the bindings from source, see the Python bindings
section of the documentation.
We aim to release MuJoCo in the first week of each month. Our versioning standards changed to modified Semantic Versioning in 3.5.0, see versioning for details.
We welcome community engagement: questions, requests for help, bug reports and feature requests. To read more about bug reports, feature requests and more ambitious contributions, please see our contributors guide and style guide.
Questions and requests for help are welcome as a GitHub "Asking for Help" Discussion and should focus on a specific problem or question.
GitHub Issues are reserved for bug reports, feature requests and other development-related subjects.
MuJoCo is the backbone for numerous environment packages. Below we list several bindings and converters.
These packages give users of various languages access to MuJoCo functionality:
- Python bindings
- dm_control, Google DeepMind's related environment stack, includes PyMJCF, a module for procedural manipulation of MuJoCo models.
- JavaScript bindings and WebAssembly support (inspired stillonearth and zalo's community projects).
- C# bindings and Unity plug-in
- MATLAB Simulink: Simulink Blockset for MuJoCo Simulator by Manoj Velmurugan.
- Swift: swift-mujoco
- Java: mujoco-java
- Julia: MuJoCo.jl
- Rust: MuJoCo-rs
- OpenSim: MyoConverter converts OpenSim models to MJCF.
- SDFormat: gz-mujoco is a two-way SDFormat <-> MJCF conversion tool.
- OBJ: obj2mjcf a script for converting composite OBJ files into a loadable MJCF model.
- onshape: Onshape to Robot Converts onshape CAD assemblies to MJCF.
If you use MuJoCo for published research, please cite:
@inproceedings{todorov2012mujoco,
title={MuJoCo: A physics engine for model-based control},
author={Todorov, Emanuel and Erez, Tom and Tassa, Yuval},
booktitle={2012 IEEE/RSJ International Conference on Intelligent Robots and Systems},
pages={5026--5033},
year={2012},
organization={IEEE},
doi={10.1109/IROS.2012.6386109}
}
Copyright 2021 DeepMind Technologies Limited.
Box collision code (engine_collision_box.c)
is Copyright 2016 Svetoslav Kolev.
ReStructuredText documents, images, and videos in the doc directory are made
available under the terms of the Creative Commons Attribution 4.0 (CC BY 4.0)
license. You may obtain a copy of the License at
https://creativecommons.org/licenses/by/4.0/legalcode.
Source code is licensed under the Apache License, Version 2.0. You may obtain a copy of the License at https://www.apache.org/licenses/LICENSE-2.0.
This is not an officially supported Google product.