MuJoCo Warp (MJWarp)

MJWarp is a GPU-optimized version of the MuJoCo physics simulator, designed for NVIDIA hardware.

MJWarp uses NVIDIA Warp to circumvent many of the sharp bits in MuJoCo MJX. MJWarp is integrated into both MJX and Newton.

MJWarp is maintained by Google DeepMind and NVIDIA.

Getting started

There are a few ways to jump into using MuJoCo Warp:

• For a quick overview of MJWarp's API and design, please see our colab that introduces the basics.
• For more details and advanced topics on using MJWarp, see the MuJoCo Warp documentation.

If you would like to train robot policies using MJWarp, consider using a robotics research toolkit that integrates it:

• MuJoCo Playground integrates MJWarp via MJX
• Isaac Lab integrates MJWarp via Newton
• mjlab integrates MJWarp directly

Installing

From PyPI:

pip install mujoco-warp

From source:

git clone https://github.com/google-deepmind/mujoco_warp.git
cd mujoco_warp
uv sync --all-extras

To make sure everything is working:

uv run pytest -n 8

If you plan to write Warp kernels for MJWarp, please use the kernel_analyzer vscode plugin located in contrib/kernel_analyzer. Please see the README there for details on how to install it and use it. The same kernel analyzer will be run on any PR you open, so it's important to fix any issues it reports.

Compatibility

The following features are implemented:

Category	Feature
Dynamics	Forward, Inverse
Transmission	All
Actuator	All except PLUGIN
Geom	All
Constraint	All
Equality	All
Integrator	All except IMPLICIT
Cone	All
Condim	All
Solver	All except PGS, noslip
Fluid Model	All
Tendon Wrap	All
Sensors	All except PLUGIN
Flex	All except flex-flex collisions, selfcollide, mjEQ_FLEXVERT, and mjEQ_FLEXSTRAIN
Mass matrix format	Sparse and Dense
Jacobian format	DENSE only (row-sparse, no islanding yet)

Differentiability via Warp is not currently available.

Viewing simulations

Explore MuJoCo Warp simulations using an interactive viewer:

mjwarp-viewer benchmarks/humanoid/humanoid.xml

This will open a window on your local machine that uses the MuJoCo native visualizer.

Batch Rendering

MJWarp includes a high-throughput GPU batch renderer designed for simultaneous rendering of cameras across many parallel simulation worlds. The renderer uses ray-tracing to render MuJoCo primitives using Warp's BVH API.

Key capabilities:

• Mesh rendering
• Texture support
• Heightfield rendering
• Flex deformable rendering
• Heterogeneous multi-camera support (different resolutions/FOV/intrinsics for each camera)
• Lighting and shadow support

Benchmarking

Benchmark as follows:

mjwarp-testspeed benchmarks/humanoid/humanoid.xml

To get a full trace of the physics steps (e.g. timings of the subcomponents) run the following:

mjwarp-testspeed benchmarks/humanoid/humanoid.xml --event_trace=True

mjwarp-testspeed has many configuration options, see mjwarp-testspeed --help for details.

Benchmark rendering with:

mjwarp-testspeed benchmarks/primitives.xml --function=render