MuJoCo (Multi-Joint dynamics with Contact) is a general-purpose physics engine designed for fast, accurate simulation of articulated systems — particularly robots and biomechanical models. Originally developed by Roboti LLC, it was acquired by Google DeepMind in October 2021 and open-sourced under the Apache 2.0 licence in May 2022. MuJoCo combines generalised (joint) coordinates — efficient for chains of bodies — with a modern optimisation-based contact solver, giving it both numerical accuracy and the ability to handle stiff constraints without the tiny time-steps required by spring-damper approaches. Written in pure C with no dynamic memory allocation, it is highly portable and deterministic, with Python bindings widely used in research. Typical applications include reinforcement-learning policy training (it powers the DeepMind Control Suite and many Gymnasium environments), system identification, model-based control, motion planning, and biomechanics research. The companion MJX library brings GPU-accelerated batched simulation via JAX. MuJoCo’s combination of speed, accuracy, and openness has made it the default simulator for academic robotics research, including DeepMind’s locomotion work and Boston Dynamics’s Spot RL training pipelines.
Open-source physics engine ('Multi-Joint dynamics with Contact') maintained by Google DeepMind. Provides fast, accurate simulation of articulated robots and contact dynamics in generalised coordinates. The standard engine for reinforcement-learning research in locomotion and manipulation.
MuJoCo (Multi-Joint dynamics with Contact) is a general-purpose physics engine designed for fast, accurate simulation of articulated systems — particularly robots and biomechanical models. Originally developed by Roboti LLC, it was acquired by Google DeepMind in October 2021 and open-sourced under the Apache 2.0 licence in May 2022. MuJoCo combines generalised (joint) coordinates — efficient for chains of bodies — with a modern optimisation-based contact solver, giving it both numerical accuracy and the ability to handle stiff constraints without the tiny time-steps required by spring-damper approaches. Written in pure C with no dynamic memory allocation, it is highly portable and deterministic, with Python bindings widely used in research. Typical applications include reinforcement-learning policy training (it powers the DeepMind Control Suite and many Gymnasium environments), system identification, model-based control, motion planning, and biomechanics research. The companion MJX library brings GPU-accelerated batched simulation via JAX. MuJoCo’s combination of speed, accuracy, and openness has made it the default simulator for academic robotics research, including DeepMind’s locomotion work and Boston Dynamics’s Spot RL training pipelines.