mujoco-warp

What it is and what it does

MuJoCo Warp is a GPU-accelerated physics simulator maintained by Google DeepMind and NVIDIA, built on top of the MuJoCo engine and NVIDIA's Warp compute framework. It brings high-throughput simulation to robotics research by offloading physics computation to NVIDIA GPUs, supporting rigid bodies, contacts, soft bodies, cloth, and signed distance fields. The simulator is designed as a near drop-in replacement for MuJoCo, with the same API surface minus a few unsupported features (IMPLICITFAST integrator, PGS solver, PLUGIN actuators, and experimental Flex support).

The package includes a high-throughput GPU batch renderer capable of rendering millions of frames per second across many parallel simulation worlds, with support for meshes, textures, heightfields, deformable bodies, heterogeneous multi-camera setups, and lighting. It integrates with JAX via MJX and with PyTorch through Isaac Lab and mjlab, making it suitable for robotics machine learning workflows. Installation is straightforward via pip, though an NVIDIA GPU is essential for production use; CPU mode is available for development and debugging.

Use it for:

• Train reinforcement learning policies for robotic control by running many parallel simulations on GPU.
• Render training datasets for vision-based robotics by batch-rendering multiple camera views across simulation worlds.
• Prototype robot behaviors and physics interactions during development using CPU mode before scaling to GPU.
• Integrate differentiable physics into JAX-based machine learning pipelines via MJX.
• Benchmark and profile physics simulation performance using the included mjwarp-testspeed tool with event tracing.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

GPU-accelerated physics simulation for robotics using NVIDIA Warp, providing high-throughput MuJoCo-compatible simulation with batch rendering across parallel worlds.

Yes, if you have an NVIDIA GPU and need high-throughput physics simulation for robotics research. The package is actively maintained, has no known vulnerabilities, uses a permissive license, and offers a GPU-accelerated alternative to standard MuJoCo with batch rendering. Start with CPU mode for development if you lack GPU access. Not suitable for CPU-only environments requiring production performance.

Install

mujoco-warp on PyPI

pip install mujoco-warp

uv add mujoco-warp

poetry add mujoco-warp

Installing mujoco-warp

Before you install

Low friction install via pip; actively maintained with recent releases (17 days old). Requires NVIDIA GPU for production use but supports CPU for development.

License in practice

Apache-2.0 permissive license allows commercial and private use with minimal restrictions.

Quickstart

pip install mujoco-warp

import mujoco_warp as mjw
# Load and simulate a MuJoCo model on GPU

NVIDIA GPU strongly recommended for performance; CPU mode available for development but not production.

Verify before relying

• Whether IMPLICITFAST integrator, PGS/noslip solvers, and PLUGIN actuator types will be supported in future releases.
• Current state of differentiability support via Warp (issue #500 referenced in description).
• Performance characteristics and throughput numbers on specific NVIDIA GPU architectures.
• Compatibility with PyTorch and JAX ecosystems beyond the mentioned integration paths.

Package facts

Evidence: mujoco_warp-3.11.0-py3-none-any.whl

Tags

More Scientific/Engineering packages

Further reading