Back to Gap Map Model components / Robotics & embodied AI

MuJoCo Playground

Google
open source / Overall score: 2.8

MuJoCo Playground is Google DeepMind's suite of robot environments, reinforcement learning tasks and hardware workflows built on MuJoCo and MuJoCo Warp. It covers large-batch RL training on the GPU with state and camera observations, teleoperation, system identification against logged hardware trajectories, and zero-shot deployment of the same policy to real robots such as the SO-101 arm.

Openness

5 high confidence
5.0
license
Apache-2.0 (OSI, covers all repository content)
source
public
core features withheld
no — Google sells no edition of it

Everything in the repository is Apache-2.0 and the suite builds from it in full. Google publishes it as a research project rather than a supported product and sells nothing alongside it.

Adoption

1 medium confidence
1.0

PyPI downloads of playground, the package the README installs. Runs through uvx and installs from a source checkout fetch the same package or skip it, so the count does not separate one-off trials from regular use.

Capability

4 medium confidence
4.0

MuJoCo Playground trains thousands of environments in parallel on the GPU from state or camera observations and carries the same policy onto real hardware. Its physics all comes from MuJoCo, where Isaac Lab lets a task switch between several physics backends, so it sits a step below Isaac Lab.

  • https://raw.githubusercontent.com/google-deepmind/mujoco_playground/main/README.md recorded 2026-11-12

    README: "built on reaf-sim, MuJoCo, and MuJoCo Warp"; "Large-batch RL training on GPU: with state and vision observations across control, manipulation, and humanoid whole-body tracking"; "System identification"; "Zero-shot sim-to-real policy deployment"; "Train SO101PickCube (8,192 parallel envs on GPU)"; "train the wrist-camera policy (SO101PickCubeVision)".

Verified 2026-11-12, except some scores, still awaiting re-confirmation