mjlab example: Unitree G1 Double Spin Kick

An example of building on top of mjlab to teach a Unitree G1 humanoid to perform a double spin kick. The reference motion comes from Jason Peng's MimicKit.

This repository provides:

• Data conversion script (MimicKit pkl → mjlab csv)
• Full training results and hyperparameters
• Pretrained ONNX checkpoint for deployment
• Deployment instructions for real hardware

⚠️ Disclaimer

This repository is provided as-is for educational purposes. We do not take any responsibility for damage to property or injury to persons that may occur from attempting to replicate the results shown here. Please exercise caution and good judgement when working with hardware.

Installation

1. Clone mjlab:

git clone https://github.com/mujocolab/mjlab

2. Clone this repo:

git clone https://github.com/mujocolab/g1_spinkick_example.git
cd g1_spinkick_example

3. Install mjlab as an editable package:

uv add --editable ../mjlab

Data Conversion

Download the spin kick data using the link referenced in the MimicKit installation instructions.

The conversion script adds smooth transitions from a safe standing pose at the start and end of the motion to ensure safe deployment. Since the motion is cyclic, we also repeat it to reach the desired duration.

# Convert pkl to csv
uv run pkl_to_csv.py \
    --pkl-file g1_spinkick.pkl \
    --csv-file g1_spinkick.csv \
    --duration 2.65 \
    --add-start-transition \
    --add-end-transition \
    --transition-duration 0.5 \
    --pad-duration 1.0

# Convert csv to npz
MUJOCO_GL=egl uv run -m mjlab.scripts.csv_to_npz \
    --input-file g1_spinkick.csv \
    --output-name mimickit_spinkick_safe \
    --input-fps 60 \
    --output-fps 50 \
    --render

This will upload the motion to your wandb registry and create a video showing the reference motion like the one below:

Training

Follow the registry creation instructions in the mjlab README.

MUJOCO_GL=egl CUDA_VISIBLE_DEVICES=0 uv run train.py \
    Mjlab-Spinkick-Unitree-G1 \
    --registry-name {your-organization}/{registry-name}/mimickit_spinkick_safe \
    --env.scene.num-envs 4096 \
    --agent.max-iterations 20_000

For full training details and reproducibility, see the wandb report.

Evaluation

To evaluate your trained policy, you'll need your wandb run path. You can find this in the run overview. It follows the format {your-organization}/{project-name}/{run-id}, where run-id is a unique 8-character identifier.

Once you have your run path, evaluate the policy with:

uv run play.py \
    Mjlab-Spinkick-Unitree-G1-Play \
    --wandb-run-path {wandb-run-path} \
    --num-envs 8

Deployment

We use motion_tracking_controller to deploy the trained policy. An ONNX file is provided for convenience, though one will also be generated in your wandb artifacts. Download it and follow the instructions in the motion_tracking_controller repo.

Alternative Implementation

An alternative option is available through RoboJuDo, a lightweight deployment framework designed for onboard execution and flexible modular integration.

RoboJuDo supports this mjlab example via the BeyondmimicPolicy, allowing direct use of the provided configuration.

A minimal configuration setup is shown below:

1. Download the ONNX model and place it in assets/models/g1/beyondmimic.

2. Edit the g1_beyondmimic config in robojudo/config/g1/g1_cfg.py to set policy name:

@cfg_registry.register
class g1_beyondmimic(RlPipelineCfg):
    robot: str = "g1"
    env: G1MujocoEnvCfg = G1MujocoEnvCfg()
    ctrl: list[KeyboardCtrlCfg] = [KeyboardCtrlCfg()]

    policy: G1BeyondMimicPolicyCfg = G1BeyondMimicPolicyCfg(
        policy_name="spinkick_safe",
        max_timestep=183,
    )

3. For real robot deployment, you may use the g1_locomimic_beyondmimic or g1_locomimic_beyondmimic_real config in robojudo/config/g1/g1_loco_mimic_cfg.py, which includes smooth locomotion interpolation.