google-deepmind/reaf-sim

google-deepmind/reaf-sim

Description: simulation framework for real robot workflows

Language: Python

License: Apache-2.0

Stars: 2

Forks: 1

Open issues: 0

Created: 2026-07-28T23:19:02Z

Pushed: 2026-10-19T23:55:08Z

Default branch: main

Fork: no

Archived: no

README:

reaf-sim

 

reaf-sim is a simulation environment framework written in PyTorch and built on MuJoCo and MuJoCo Warp with the Robotics Environment Authoring Framework (REAF).

reaf-sim enables creation of environments in simulation that can be zero-shot deployed onto hardware and vice versa. Workflows that reaf-sim enables include large-batch RL on GPU, sim evaluation, sim and hardware teleoperation, system identification, agentic workflows, and sim2real deployment.

This project adheres to the GDM Robotics Environment interface.

> Looking for pre-built robot environments and training scripts? We will be revamping MuJoCo Playground with ready-made examples for RL, sim/real teleop, sysid, sim2real, and agentic robotics. Stay tuned!

Quickstart

Clone reaf-sim from GitHub and install with `uv`:

git clone https://github.com/google-deepmind/reaf-sim.git && cd reaf-sim
make install # install dependencies (use `make install-cpu` for CPU only)
make test # run all tests, verify everything works

Cartpole Example

For a quick preview, let's create a reaf-sim environment controlling a cartpole system alongside a camera rendering RGB observations across 16 parallel environments on GPU (impl="warp"). Step the environments with control commands and capture visual rollouts:

import mujoco
from reaf.simulation.torch import ActuatorDevice, CameraDevice, flat_scene
import torch

spec = mujoco.MjSpec.from_file("cartpole.xml")

# configure the environment
config = flat_scene.Environment(
base_scene=spec,
sim=flat_scene.Sim(nworld=16, xpu_device="cuda:0", impl="warp"),
device_layer=flat_scene.DeviceLayer(
devices={
"slider": flat_scene.Component(
ActuatorDevice, targets=["slider"]
).configure_spec(mode="motor"),
"camera": flat_scene.Component(
CameraDevice,
cam_name="fixed",
cam_resolution=(160, 120),
render_rate_hz=20.0,
output_name_prefix="camera/",
),
}
),
)

# build the environment
env = flat_scene.build_env(config)

# step the environment with batched actions and extract camera frames
timestep = env.reset()
for _ in range(100):
commands = torch.rand((env.nworld, 1), device="cuda:0") * 2.0 - 1.0
timestep = env.step({"slider/ctrl": commands})
camera_frames = timestep.observation["camera/fixed_rgb"] # (16, 120, 160, 3)

For the complete, runnable script that loads the Cartpole model and generates the tiled grid video below, see [reaf/simulation/torch/examples/cartpole.py](reaf/simulation/torch/examples/cartpole.py).

Documentation & Contributing

• Documentation: See the reaf-sim documentation for the user guide, architecture overview, and API reference.
• Contributing: See [CONTRIBUTING.md](CONTRIBUTING.md) for contribution guidelines.
• Changelog: See [docs/changelog.rst](docs/changelog.rst).

Citation

If you use reaf-sim in your work, please cite it as follows (see the GitHub contributors page for the full list of contributors):

@software{reaf_sim_2026,
author = {{reaf-sim Contributors}},
title = {{reaf-sim: Robotics Environment Authoring Framework - Simulation}},
url = {https://github.com/google-deepmind/reaf-sim},
year = {2026},
}

Acknowledgements

We are grateful to all members of the MuJoCo team for their support in developing reaf-sim. Additionally, we thank the authors of REAF, mjlab, and Isaac Lab for inspiring reaf-sim.

License & Disclaimer

reaf-sim is released under the Apache 2.0 license. See [LICENSE](LICENSE) for details.

This is not an official Google product.