MuJoCo Playground

A collection of GPU-accelerated robot learning environments led by DeepMind, built for fast single-GPU training and zero-shot transfer to real robots.

MuJoCo Playground is a collection of robot learning environments open-sourced in February 2025 under the Apache 2.0 license, led by Google DeepMind together with UC Berkeley and other teams. Physics runs on MJX, can now also be switched to the MuJoCo Warp backend, and it ships with a batched renderer for training policies that take images as input. It bundles together the classic tasks from the DeepMind Control Suite along with legged-locomotion and manipulation tasks, covering robots such as Unitree's Go1, G1, and H1, the Booster T1, the Berkeley Humanoid, Spot, Franka, ALOHA, and the LEAP dexterous hand. The paper reports training policies in minutes on a single GPU and demonstrates zero-shot transfer to real robots from both state and pixel inputs, cutting out a large amount of the usual repetitive work of building environments and tuning rewards. It installs with pip install playground.

ExampleAfter installing it, a user can load the G1 or Go1 walking environment, train thousands of simulated robots in parallel on one GPU with PPO, and then deploy the trained policy straight to the real robot for testing.

Also called

Playground

Related

MuJoCo XLA · MuJoCo Warp · MuJoCo (Multi-Joint dynamics with Contact) · DeepMind Control Suite · Sim-to-Real Transfer · mjlab

Sources

MuJoCo Playground (arXiv 2502.08844)
google-deepmind/mujoco_playground (GitHub)
MuJoCo Playground 项目主页 (Chinese)

As of

2026-09