MuJoCo Playground
Google DeepMind's MuJoCo suite for training robots in sim and running them on real arms
License
- code (the whole repository, including its copy of reaf-sim): Apache 2.0
- robot models (MuJoCo Menagerie: SO-101; YAM; WidowX AI and Unitree G1): Apache 2.0, MIT, BSD 3-Clause
Code: Apache-2.0, per LICENSE, which covers the whole repository including its copy of reaf-sim. Robot models load from MuJoCo Menagerie under their own permissive licences: SO-101 Apache-2.0, YAM MIT, WidowX AI and Unitree G1 BSD-3-Clause. The G1 tracking example uses motion clips you download separately; the LAFAN1 motion capture they are retargeted from is CC BY-NC-ND 4.0, so check that data before commercial use. Checked 2026-11-09 against LICENSE and the Menagerie model licences.
Overview
MuJoCo Playground is an open-source suite of robot environments and hardware workflows from Google DeepMind's robotics simulation team, built on MuJoCo, MuJoCo Warp and the reaf-sim environment framework. Version 1.0, released on 6 November 2026, rebuilds it in PyTorch and pairs each robot's simulation model with a hardware driver behind the same interface, so a task and a trained policy run in simulation or on the real robot by changing one flag. It ships large-batch reinforcement learning with RSL-RL, leader-follower teleoperation, system identification, zero-shot sim-to-real deployment and a harness in which a vision-language model controls a tabletop arm. The SO-101 and i2rt YAM Pro arms are supported in simulation and on hardware, the Trossen WidowX AI arm and the Unitree G1 humanoid in simulation. The code is licensed Apache-2.0.
MuJoCo Playground is a suite of robot environments, reinforcement learning tasks and hardware workflows built on MuJoCo, MuJoCo Warp and reaf-sim, the simulation side of Google DeepMind's Robotics Environment Authoring Framework. Every environment is split into two layers. A task layer turns actions into commands and computes observations, rewards and terminations, and it never sees the simulator's privileged state. A device layer takes those commands and returns measurements, from MuJoCo on the CPU, MuJoCo Warp on the GPU or a physical robot. Because the task layer only talks to the device layer, the same task code and the same trained policy run in simulation (--mode=sim) or on hardware (--mode=real); the reaf-sim guide explains the split.
Five workflows come with it. Reinforcement learning trains PPO policies with RSL-RL across thousands of parallel simulated worlds on a GPU, from state or camera pixels, for classic control, manipulation and humanoid whole-body motion tracking; the training guide has the details, and the quickstart trains a cart-pole balance policy in about 1 minute on a CPU. Teleoperation drives a follower arm from a physical leader arm, in simulation or on hardware, with a three-button foot pedal to start and score episodes and MCAP logs that replay in Rerun, as the teleoperation guide shows. System identification uses MuJoCo's own sysid toolbox to fit joint armature, friction, stiffness and damping to trajectories logged on the real arm. Sim-to-real deployment runs a wrist-camera policy that picks up a cube, trained in simulation with domain randomization, on the physical SO-101 without further training, while a live digital twin in MuJoCo Studio mirrors the arm in the browser. The agentic harness lets a vision-language model control an arm turn by turn: a local bridge serves the robot state and an annotated wrist-camera image, and the model replies with a Python plan built from inverse kinematics primitives such as look_at, grasp and place.
Each robot has its own package with a simulation model and, where hardware is supported, a driver with the same joint order and control interface. The SO-101 and the i2rt YAM Pro arms run in simulation and on hardware, and the SO-101 hardware extra installs LeRobot for its Feetech servos; the Trossen WidowX AI arm and the Unitree G1 humanoid are simulation only. The robot models are loaded from MuJoCo Menagerie, and the SO-101 and YAM come with dynamics already identified on real hardware. The G1 example tracks motion clips in the style of BeyondMimic. Hardware drivers are optional extras, so a simulation-only install stays light. The package is called playground on PyPI, needs Python 3.12 or 3.13, and its examples run straight from uvx on a CPU or on NVIDIA GPUs with CUDA 12 or 13.
Version 1.0 was released on 6 November 2026 and is a rebuild rather than an update. The earlier version, the RSS 2025 research preview described in a 2025 technical report, was a set of GPU reinforcement learning environments written in JAX on MJX and stayed in simulation. Version 1.0 moves to PyTorch, runs on MuJoCo or MuJoCo Warp, adds the paired simulation and hardware packages, deployment code, teleoperation, system identification and the agentic harness, and renders in MuJoCo Studio. The JAX and MJX environments are still in the repository's rss2025 folder and on the RSS 2025 archive page. Version 1.0.1 followed on 8 November 2026 with fixes, listed in the changelog. The 1.0 release is credited to Baruch Tabanpour, Taylor Howell, Sam Haves, Saminda Abeyruwan, Tony Ma, Dan Suh and the project's contributors, and the launch video shows each workflow on real arms.
Why it matters
Most simulation frameworks stop at a trained policy. Getting that policy onto a real robot means a second stack: drivers, calibration, a deployment loop and a way to compare the real arm with its model. MuJoCo Playground 1.0 ships that second half for two arms, the SO-101, the low-cost open arm widely used with LeRobot, and the i2rt YAM Pro, and for both the same task code covers training, teleoperated data collection, system identification and deployment. For readers of our MuJoCo page it is a path from MuJoCo Warp training to hardware maintained by Google DeepMind, which also develops MuJoCo. For mjlab and Isaac Lab users it is another GPU training stack, with fewer tasks but with deployment built in. SO-101 owners get a simulation model with identified dynamics and a working sim-to-real example for the arm they already have. The hardware path is days old and covers two arms, and the promise of no switching cost between simulation and hardware is the team's own; so far the one learned policy shown moving across is the SO-101 cube pick-up.
Problem it solves
Training a robot policy in simulation and running it on the real robot usually means two codebases: the simulator's environment, and a separate stack of drivers, calibration and deployment code that has to reproduce the same observations and actions. MuJoCo Playground keeps the task logic behind a device layer that is either the simulator or the robot, and ships the drivers, teleoperation, system identification and deployment scripts for supported arms, so the same task and policy move between simulation and hardware with one flag.
Key takeaways
- Google DeepMind's suite of robot environments and hardware workflows on MuJoCo and MuJoCo Warp, rebuilt in PyTorch for version 1.0
- The same task code and trained policy run in simulation or on the real robot, because tasks only see a device layer
- Workflows for GPU reinforcement learning with RSL-RL, teleoperation, system identification, sim-to-real deployment and control by a vision-language model
- SO-101 and i2rt YAM Pro arms in simulation and on hardware; Trossen WidowX AI and Unitree G1 in simulation
- Robots and policies can be watched live in MuJoCo Studio in the browser
- Apache-2.0 and installable from PyPI as playground; the RSS 2025 JAX and MJX version stays in the repository
Who it is for
- SO-101 and LeRobot users who want to train in simulation and deploy on their own arm
- Labs with i2rt YAM arms that need simulation, teleoperation and system identification for them
- Reinforcement learning researchers training on GPUs with MuJoCo Warp and PyTorch
- Teams building agents in which a vision-language model controls a robot arm
Limitations
- Hardware support covers two arms, the SO-101 and the i2rt YAM Pro; the Trossen WidowX AI and the Unitree G1 run in simulation only
- Version 1.0 came out on 6 November 2026 as a rebuild, so it has had little outside use, and version 1.0.1 on 8 November 2026 already changed the device interface
- The zero-shot sim-to-real example is one task, the SO-101 picking up a cube; the other hardware examples are teleoperation, system identification and agent control
- The agentic harness runs Python plans written by the vision-language model directly on the robot, and the documentation says to use it at your own risk
- The Unitree G1 tracking example needs motion clips you fetch yourself, such as the retargeted LAFAN1 clips, and the LAFAN1 motion capture data is licensed CC BY-NC-ND 4.0
- Large-batch training needs an NVIDIA GPU with CUDA 12 or 13; on a CPU the quickstart runs 64 parallel worlds, and the viewer on macOS needs mjpython
- The README states that it is not an officially supported Google product