YAHMP: Yet Another Humanoid Motion tracking Policy

YAHMP is a humanoid general motion tracking policy for the Unitree G1 robot.

The training pipeline builds upon mjlab.

g1-yahmp.mp4

Installation

First, install uv:

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

Then install the project dependencies with:

uv sync

Finally, verify that the YAHMP environments are correctly installed and visible with:

uv run list_envs | rg YAHMP

Motion Assets

Download the retargeted OMOMO and AMASS motions used for training from this link, extract the archive, and copy the g1_omomo_amass_clean folder into assets/motions/.

The default motion configuration src/yahmp/config/g1/motion_data_cfg.yaml specifies which motions are loaded during training—modify it if you want to use different reference datasets.

Quick Checks

Smoke-test the teacher scene with a zero-action agent:

uv run play Mjlab-YAHMP-Unitree-G1 --agent zero

Training

uv run train Mjlab-YAHMP-Unitree-G1 --env.scene.num-envs 8192

Run Pre-trained Policy

uv run python -m yahmp.scripts.deploy.run_yahmp_onnx_mujoco \
  --task-id Mjlab-YAHMP-Unitree-G1 \
  --onnx-path assets/models/g1_yahmp.onnx \
  --motion-file assets/motions/g1_omomo_amass_clean/<motion-name>.npz

Utilities

Utility scripts for data conversion, ONNX export, deployment, evaluation, and workflow helpers live under src/yahmp/scripts. See the dedicated guide at src/yahmp/scripts/README.md.

Extras

In addition to Mjlab-YAHMP-Unitree-G1, this project also includes environments that together form an example Teacher–Student training pipeline:

• Mjlab-YAHMP-Teacher-Unitree-G1: privileged teacher example
• Mjlab-YAHMP-Student-RL+Action-Matching-Unitree-G1: student trained via RL + Action-Matching distillation
• Mjlab-YAHMP-Student-RL+KL-Matching-Unitree-G1: student trained via RL + KL-Matching distillation

Development

Install the dev tools with:

uv sync --group dev

Format and auto-fix the repo with:

make format