MuJoCo Warp (MJWarp)

MJWarp is a GPU-accelerated version of the MuJoCo physics simulator, designed for NVIDIA hardware. MJWarp delivers high-throughput, accurate simulation for robotics research.

MJWarp is maintained by Google DeepMind and NVIDIA as part of the Newton project.

Getting started

MuJoCo Warp requires an NVIDIA GPU for fast simulation but supports CPU for development and debugging.

Try it now: view a simulation of a dancing humanoid robot locally on your machine:

git clone https://github.com/google-deepmind/mujoco_warp.git && cd mujoco_warp
python benchmarks/run.py -f unitree_g1_flat --view

Or try out a tutorial in your browser (no local setup required).

MJWarp is also available via PyPI:

pip install mujoco-warp

Examples

MuJoCo Warp simulates many kinds of physical systems, from rigid bodies with contacts to soft bodies, cloth, signed distance fields, and more. Here are a few examples of what it can do:

   ![Unitree G1](benchmarks/unitree_g1/rollout_flat.webp) 

python benchmarks/run.py -f unitree_g1_flat --view

   ![Unitree G1 Heightfield](benchmarks/unitree_g1/rollout_hfield.webp) 

python benchmarks/run.py -f unitree_g1_hfield --view

   ![MyoArm](benchmarks/myosim/rollout.webp) 

python benchmarks/run.py -f myoarm --view

   ![ALOHA Clutter](benchmarks/aloha/rollout_clutter.webp) 

python benchmarks/run.py -f aloha_clutter --view

   ![ALOHA Pot](benchmarks/aloha/rollout_pot.webp) 

python benchmarks/run.py -f aloha_pot --view

   ![ALOHA SDF](benchmarks/aloha/rollout_sdf.webp) 

python benchmarks/run.py -f aloha_sdf --view

   ![ALOHA Cloth](benchmarks/aloha/rollout_cloth.webp) 

python benchmarks/run.py -f aloha_cloth --view

   ![Three Humanoids](benchmarks/humanoid/rollout_three_humanoids.webp) 

python benchmarks/run.py -f three_humanoids --view

   ![Cloth](benchmarks/cloth/rollout.webp) 

python benchmarks/run.py -f cloth --view

Each of these scenes is benchmarked nightly and the results are published nightly.

Tips for developers

To set up MJWarp for development:

git clone https://github.com/google-deepmind/mujoco_warp.git && cd mujoco_warp
uv sync --all-extras  # install all optional dependencies for development
uv run pre-commit install  # enables ruff, uv-lock, and kernel-analyzer checks on commit
uv run pytest -n 8  # run all tests, verify everything works

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

For performance profiling MJWarp, use the --event_trace flag on mjwarp-testspeed to get a full trace on a test scene of your choice:

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

mjwarp-testspeed has many configuration options, see mjwarp-testspeed --help for details. For more details and advanced topics on using MJWarp, see the MuJoCo Warp documentation.

Integrating MuJoCo Warp

There are many ways to use MuJoCo Warp in your projects. In many cases, you can directly install and use MJWarp as a drop-in replacement for MuJoCo.

If you prefer the JAX ecosystem, you can use MJWarp via MJX. See MuJoCo Playground for robotics machine learning recipes that use JAX and MJWarp.

If you prefer PyTorch for research, consider one of these two great options:

• Isaac Lab integrates MJWarp via Newton. This setup enables powerful, highly extensible multi-physics simulation with deep NVIDIA ecosystem integration.
• mjlab exposes the Isaac Lab manager-based API directly on top of MJWarp, providing a focused framework for robotics research with minimal dependencies and direct access to native MuJoCo data structures.

MuJoCo API Compatibility

MuJoCo Warp supports the same features as MuJoCo with the following exceptions:

• Integrator: IMPLICITFAST midpoint integrator feature is not supported
• Solver: PGS and noslip not yet supported
• Actuator / Sensors: PLUGIN types not yet supported
• Flex: experimental — not all features are implemented or optimized yet

Differentiability via Warp is not yet available. See #500 for progress.

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 scenes at millions of frames per second on NVIDIA GPUs.

Key capabilities:

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

See the announcement PR for more details.

License

MJWarp is released under the Apache 2.0 license. See LICENSE for details.