PAL Robotics in mjlab

This repository showcases the implementation of PAL Robotics' robots in mjlab.

What is mjlab?

mjlab brings the Isaac Lab API to MuJoCo Warp. It's lightweight, easy to install, and has been validated for sim-to-real transfer on the G1 and Go1 robots for RL locomotion and motion imitation. See the announcement thread for videos, or read about the motivation behind mjlab.

Installation

Install uv.

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

Clone the repository.

git clone https://github.com/pal-robotics/pal_mjlab.git 
cd pal_mjlab
uv sync

Quick Start

List available environments.

uv run list_envs --keyword pal

Test with dummy agents.

uv run play Mjlab-Velocity-Flat-Pal-Kangaroo --agent zero    # send zero actions
uv run play Mjlab-Velocity-Flat-Pal-Kangaroo --agent random  # send random actions

Velocity Tracking

Train a locomotion policy.

uv run train Mjlab-Velocity-Flat-Pal-Kangaroo --env.scene.num-envs 4096

Evaluate a trained policy.

uv run play Mjlab-Velocity-Flat-Pal-Kangaroo --wandb-run-path your-org/mjlab/run-id

Motion Imitation

We added support to GMR to retarget animations from the LaFAN1 dataset for PAL robots (KANGAROO and Talos).

Retargeting a new motion

First, use GMR to retarget and convert a motion file. Note that for convenience we provide a few retargeted motions under motions folder.

Then convert the CSV to NPZ format.

uv run -m pal_mjlab.scripts.csv_to_npz \
    --input-file motions/kang_walk.csv \
    --output-name kang_walk \
    --input-fps 30 \
    --output-fps 50 \
    --robot-name kangaroo \
    --render True

Training and evaluation

Train.

uv run train Mjlab-Tracking-Flat-Pal-Kangaroo \
    --registry-name your-org/csv_to_npz/kang_walk \
    --env.scene.num-envs 4096

Evaluate.

uv run play Mjlab-Tracking-Flat-Pal-Kangaroo --wandb-run-path your-org/mjlab/run-id

Results

pal_mjlab.mp4

Contributing

Contributions are welcome! Please open an issue to discuss proposed changes or report bugs. As mjlab is in early development, breaking changes may occur—thank you for your patience.

Acknowledgements

Thanks to the teams behind mjlab, PAL Robotics, MuJoCo Warp, Isaac Lab, and the Inria HUCEBOT Team.

License

See LICENSE for details.