Key takeaways
- A free physics engine from Google DeepMind simulates joints and contact for robotics and machine learning research.
- Try it with nothing to install: the project offers a browser version on desktop or mobile.
- Work in Python, C, Unity or the web: the project provides bindings and a plug-in beyond its core C API.
- Explore differentiable physics: one tutorial trains locomotion policies with gradients derived from the physics step.
- The catches: it targets researchers, the main branch may be unstable, and it is not an officially supported Google product.
What it is
MuJoCo, short for multi-joint dynamics with contact, is a general purpose physics engine. It is open source under the Apache licence, and the code is on GitHub at google-deepmind/mujoco.
The problem it solves
Researchers in robotics, biomechanics, graphics and animation, and machine learning need simulation that is both fast and accurate when jointed bodies touch their surroundings.
MuJoCo is aimed at that audience. It has a C API and is intended for researchers and developers. The repository is maintained by Google DeepMind.
What you get
- Speed: the runtime simulation module is tuned for performance.
- A quick look: the project's website lets you explore MuJoCo on desktop or mobile with nothing to install.
- Flexibility: Python bindings and a Unity plug-in are provided alongside the C API.
- Machine learning routes: MJX is a branch of MuJoCo written in JAX.
- Differentiable physics: a tutorial trains locomotion policies using gradients derived from the physics step.
The catches
- Audience: this is a library for researchers and developers, not a finished consumer app.
- Browser limits: the WebAssembly version is slower than native C, by the project's own description.
- Maturity: the tip of the main branch may be unstable, so the project recommends its prebuilt releases, which it aims to publish monthly.
- Platforms: prebuilt Windows binaries cover only one processor family.
- Support: the README says this is not an officially supported Google product.
What we did not test
We did not run MuJoCo, try the browser version, build any model or measure its speed. Every claim here comes from a literal quote in the project's README or the GitHub API, and the feature descriptions are the project's own.
Verdict
Editors' view: this is for people building or studying simulated robots, bodies or learning agents who want a permissively licensed engine from a well-known lab. If you only want a game physics toy, it is probably more than you need. Curious readers can start with the browser demo before committing to anything.
Sources
Need help adopting this, or have a software project in mind?
Tell us what you are building. The initial consultation is free. Email [email protected] or follow @zolnaltd.
Start a conversation →