MuJoCo
In plain wordsA free physics simulator, now maintained by Google DeepMind, that AI researchers use to train and test simulated robots and bodies, especially with trial-and-error learning.
An open-source physics engine (Multi-Joint dynamics with Contact) designed for model-based control, robotics and biomechanics, and a standard tool in reinforcement learningreinforcement learning, n.: A way of teaching a computer program by trial and error: it tries actions, earns points for good results and penalties for bad ones, and gradually learns which choices pay off over time. research. It represents systems in generalized (joint) coordinates, which keeps articulated bodies such as arms, hands and legged robots stable at relatively large time steps, and it models contact through a convex optimization that permits soft, slightly penetrating contacts with tunable parameters, trading some physical realism for speed, smoothness and derivatives useful in trajectory optimizationtrajectory optimization, n.: Calculating the best way for a robot to move over time, choosing exact positions, speeds and forces moment by moment so the motion is smooth, fast or energy-efficient and avoids obstacles. and model predictive controlmodel predictive control, n.: A control method that repeatedly plans the best moves a few seconds ahead using a model of the machine, carries out only the first move, then plans again with fresh measurements.. Models are written in the MJCF XML format. MuJoCo underlies widely used benchmarks — the continuous-control locomotion tasks of OpenAI Gym and its successor GymnasiumGymnasium, prop. n.: A free software library that gives reinforcement learning programs one standard way to plug into practice tasks, from simple games to simulated robots, so methods can be tried on identical problems. It succeeded OpenAI's Gym., and the DeepMind Control Suite — so 'MuJoCo tasks' often denotes that family of benchmarks. MJX, a reimplementation in JAXJAX, prop. n.: A free Google library that lets researchers write mathematics in familiar Python and automatically get derivatives and very fast code for graphics chips and Google's AI chips., runs thousands of simulations in parallel on accelerators, and further GPU versions followed in 2025.
Our agent beats the baseline on all six MuJoCo tasks but has never touched a real robot.
Origin
Created by Emanuel Todorov with Tom Erez and Yuval Tassa at the University of Washington and described at IROS 2012; it was sold commercially by Todorov's company Roboti LLC until Google DeepMind acquired it in October 2021 and released it as open source under the Apache 2.0 license in May 2022.
Usage note
Results on 'MuJoCo' benchmarks depend on the library and environment versions, so scores reported years apart may not be comparable.
Related
- physics simulatorphysics simulator, n.: Software that imitates how real objects move, collide, bend and fall, so robots and their AI can be designed, trained and tested on a computer before touching the real world.
- GymnasiumGymnasium, prop. n.: A free software library that gives reinforcement learning programs one standard way to plug into practice tasks, from simple games to simulated robots, so methods can be tried on identical problems. It succeeded OpenAI's Gym.
- reinforcement learningreinforcement learning, n.: A way of teaching a computer program by trial and error: it tries actions, earns points for good results and penalties for bad ones, and gradually learns which choices pay off over time.
- sim-to-real transfersim-to-real transfer, n.: Training a robot's AI in a computer simulation, where practice is fast, cheap and safe, and then making it work on the real robot despite the differences between simulation and reality.
- model predictive controlmodel predictive control, n.: A control method that repeatedly plans the best moves a few seconds ahead using a model of the machine, carries out only the first move, then plans again with fresh measurements.
- JAXJAX, prop. n.: A free Google library that lets researchers write mathematics in familiar Python and automatically get derivatives and very fast code for graphics chips and Google's AI chips.
Sources
- Todorov, E., Erez, T. & Tassa, Y. (2012). MuJoCo: A physics engine for model-based control. Proc. IEEE/RSJ IROS 2012, 5026–5033.
Mentioned in
GymnasiumGymnasium, prop. n.: A free software library that gives reinforcement learning programs one standard way to plug into practice tasks, from simple games to simulated robots, so methods can be tried on identical problems. It succeeded OpenAI's Gym., physics simulatorphysics simulator, n.: Software that imitates how real objects move, collide, bend and fall, so robots and their AI can be designed, trained and tested on a computer before touching the real world., sim-to-real transfersim-to-real transfer, n.: Training a robot's AI in a computer simulation, where practice is fast, cheap and safe, and then making it work on the real robot despite the differences between simulation and reality.