MuJoCo Warp

A GPU rewrite of MuJoCo built by DeepMind and NVIDIA on the Warp framework, aimed at massive batched parallelism.

MuJoCo Warp is a GPU version of MuJoCo developed jointly by Google DeepMind and NVIDIA, reimplemented using NVIDIA Warp (a framework for writing GPU kernels in Python), and it also serves as the core solver of the Newton physics engine. It is built for throughput, advancing hundreds or thousands of simulated “worlds” at once; its developers say it scales better than MJX on scenes with many geometries, high degrees of freedom, and complex contact, and that it integrates more easily with PyTorch. The trade-off is that its per-step latency can be slower than plain MuJoCo, so real-time control still uses the original; it also doesn't support automatic differentiation, so differentiable simulation still requires MJX's JAX implementation. It additionally supports batched ray-traced rendering. Since January 2026 it has shipped on PyPI as mujoco-warp, with version numbers kept in sync with MuJoCo, and it can be called from mjlab, MJX, MuJoCo Playground, and Newton.

ExampleA typical workflow uses mjw.put_model to load a model onto the GPU, mjw.make_data(mjm, nworld=100) to create 100 worlds, and a single mjw.step call to advance all 100 simulations one step at once, with a CUDA Graph capturing the loop for further speedup.

Also called

MJWarp, mujoco-warp, MJX-Warp

Related

MuJoCo (Multi-Joint dynamics with Contact) · MuJoCo XLA · Newton Physics Engine · NVIDIA Warp · mjlab · GPU-Accelerated Parallel Simulation

Sources

google-deepmind/mujoco_warp (GitHub)
MuJoCo Warp documentation
mujoco-warp (PyPI)

As of

2026-09