mjlab 0.0.4

IsaacLab-style manager API, powered by MuJoCo-Warp, for RL and robotics research.

mjlab

⚠️ EXPERIMENTAL PREVIEW

This project is in very early experimental stages. APIs, features, and documentation are subject to significant changes. Use at your own risk and expect frequent breaking changes.

Isaac Lab API with MJWarp backend.

uv run demo

Development Guide

Clone mjlab:

git clone [e-mail]:mujocolab/mjlab.git && cd mjlab

Using uv

Install uv:

curl -LsSf https://astral.sh/uv/install.sh | sh

Once installed, you can verify it works by running:

uv run scripts/list_envs.py

Reinforcement Learning

Velocity training

Train a Unitree G1 to follow velocity commands (headless, large batch):

MUJOCO_GL=egl uv run train \
  Mjlab-Velocity-Flat-Unitree-G1 \
  --env.scene.num-envs 4096

Play the trained policy:

uv run play --task Mjlab-Velocity-Flat-Unitree-G1-Play

Motion mimicking

Run a pre-trained motion-mimic policy on the G1:

uv run play \
  --task Mjlab-Tracking-Flat-Unitree-G1-Play \
  --wandb-run-path gcbc_researchers/mjlab_alpha/rfdej55h

Train the same motion-mimic policy (headless, large batch):

MUJOCO_GL=egl uv run train \
  Mjlab-Tracking-Flat-Unitree-G1 \
  --registry-name gcbc_researchers/csv_to_npz/lafan_cartwheel \
  --env.scene.num-envs 4096

Add a new motion to the WandB registry from a CSV:

MUJOCO_GL=egl uv run src/mjlab/scripts/csv_to_npz.py \
  --input-file /path/to/motion.csv \
  --output-name side_kick \
  --input-fps 30 \
  --output-fps 50 \
  --render

Running tests

make test

Code formatting and linting

You can install a pre-commit hook:

uvx pre-commit install

or manually format with:

make format

Troubleshooting

CUDA Compatibility: Not all CUDA versions are supported. Check mujoco_warp#101 for your CUDA version compatibility.

License

This project, mjlab, is licensed under the Apache License, Version 2.0.

Third-Party Code

The third_party/ directory contains selected files from external projects.
Each such subdirectory includes its own original LICENSE file from the upstream source.
These files are used under the terms of their respective licenses.

Currently, third_party/ contains:

• isaaclab/ — Selected files from NVIDIA Isaac Lab,
  licensed under the BSD-3-Clause.

When distributing or modifying this project, you must comply with both:

1. The Apache-2.0 license of mjlab (applies to all original code in this repository).
2. The licenses of any code in third_party/ (applies only to the files from those projects).

See the individual LICENSE files for the complete terms.