MuJoCo

MuJoCo (Multi-Joint dynamics with Contact) is a general-purpose physics simulator developed by Google DeepMind (originally by Roboti LLC, acquired by DeepMind in 2021). It is one of the most widely used physics engines in robotics and embodied AI research, with over 15,000 GitHub stars and licensed under Apache 2.0. MuJoCo was designed from the ground up for model-based optimization and simulation of multi-joint dynamics with frictional contact. Its key innovations include a fast and accurate contact model using convex optimization, support for arbitrary 3D geometries including meshes, and efficient computation of dynamics through recursive Newton-Euler algorithms. The simulator supports a wide range of applications including robot control, biomechanics, animation, and reinforcement learning. It has been used extensively to train locomotion and manipulation policies that transfer to real robots, and is the default physics engine in numerous robot learning frameworks including DeepMind's own research stack. In 2022, DeepMind open-sourced MuJoCo under the Apache 2.0 license, making it freely available for both research and commercial use. Since then, it has gained Python bindings, support for differentiable physics, and integration with popular RL libraries. MuJoCo's combination of speed, accuracy, and ease of use makes it the preferred physics simulator for many embodied AI projects. It is used by both the Isaac Lab and Habitat simulators as an underlying physics engine, and is the default simulator for the OpenAI Gym robotics environments. Version 3.13.0 of MuJoCo was released on September 9, 2026 (release notes dated September 8, 2026). The release allows site geometries to be associated with meshes, adds an insidesite sensor together with the mj_insideSite function, introduces a new discrete integrator that merges the constraint solve and the implicit velocity update into a single operation, adds support for Python 3.15 including free-threading, and changes the .mjz archive encoder to write the root file as model.xml.