Skip to main content

Listed

, home
FREN

MuJoCo

Definition

A general-purpose physics engine, whose name stands for Multi-Joint dynamics with Contact, designed to simulate articulated structures in contact with their environment quickly and accurately. It targets robotics, biomechanics, animation and machine learning. Originally developed by Roboti LLC, it was bought by Google DeepMind, which made it free in 2021 and then open source in 2022.

Checked on against the references below. · Written with AI · checked against the sources Source: MuJoCo documentation, “Overview”.

Concrete examples

For the Unitree Go2 and the Unitree G1, Unitree publishes unitree_mujoco, a simulator “developed based on Unitree sdk2 and mujoco”, mainly used to check a controller before moving to the real robot. LimX Dynamics’ tron1-mujoco-sim repository lets users run the official controllers and RL policies of the TRON 1 in MuJoCo, following a “Sim-to-sim before sim-to-real” approach. For the Stretch 4, Hello Robot publishes stretch4_mujoco, a simulation stack built on MuJoCo.

What it changes for you

MuJoCo Warp requires an NVIDIA GPU for fast simulation, while CPU remains supported for development and debugging. mjlab, which builds on MuJoCo Warp, also requires an NVIDIA GPU for training. On a single scene, the MJX documentation warns that MJX-JAX can be ten times slower than MuJoCo, which has been carefully optimized for CPU.

Not to be confused with

MJX is a JAX API that currently supports two implementations of MuJoCo, MJX-JAX and MJX-Warp, the latter built on MuJoCo Warp, designed for hardware accelerators. MuJoCo Warp supports MuJoCo’s features with some exceptions, including the PGS solver, not yet supported. MJX-Warp does not support automatic differentiation, unlike MJX-JAX.

Additional material checked on .

Related terms

References

Back to the glossary