H-ACT is the action execution module within the FRoM-W1 framework. It converts the action representation sequence generated by H-GPT into executable motions on humanoid robots. This module contains three key components of FRoM-W1:

• Retarget
  Converts H-GPT action representations into SMPLX motion sequences and further retargets them to various humanoid robots and dexterous hands. This also enables retargeting human motion datasets such as AMASS to multiple robot platforms.

• Policy Training
  Trains action execution policies based on retargeted motion data (or other robot datasets).

• Sim2sim & Sim2real
  Deploys trained policies to simulation or real robots, enabling sim2sim and sim2real execution.

Motion Reconstruction & Retargeting

This part is responsible for reconstructing the 623-dimensional action representation generated by H-GPT into SMPLX motion sequences, and retargeting them to the joint space of target humanoid robots and dexterous hands.

🧩 Environment Setup & Model Preparation

Enter the H-ACT/retarget directory and configure the retarget environment:

conda create -n retarget python=3.10
conda activate retarget
pip install -r requirements.txt

This module requires SMPL and MANO models. Before use, download the corresponding model files:

1. Download MANO models Visit the MANO official website and download the model files (MANO_LEFT.pkl, MANO_RIGHT.pkl) into models/mano.

2. Download SMPL models Visit the SMPL official website and download (SMPL_NEUTRAL.pkl, SMPL_MALE.pkl, SMPL_FEMALE.pkl) into models/smpl.

Example directory structure:

retarget
├── models
│   ├── mano
│   │   ├── MANO_LEFT.pkl
│   │   └── MANO_RIGHT.pkl
│   └── smpl
│       ├── SMPL_NEUTRAL.pkl
│       ├── SMPL_MALE.pkl
│       └── SMPL_FEMALE.pkl
├── ...

📁 Data Preparation

Create the data directory inside retarget to store input and output data:

• data/623: stores the 623-dimensional action data from H-GPT
• data/smplx: stores intermediate SMPLX motion sequences
• data/output: stores final robot & dexterous-hand joint sequences

Example structure:

retarget
├── data
│   ├── 623
│   │   ├── data1.npy  # Output from H-GPT
│   │   └── data2.npy
│   ├── smplx          # Output SMPLX directory
│   ├── output         # Output robot motion directory
├── ...

▶️ Run (Retarget Execution)

Run the following command to retarget the action representations into robot-specific joint sequences:

python main.py

The module currently supports the following robots and dexterous hands:

• Unitree H1
• Unitree G1
• Dex3
• InspireHand

You can modify lines 47–48 in main.py to select a target robot.

Policy Training

In this stage, we train an action execution policy based on the motion data generated during the Retarget step (or any other robot motion dataset).
Building upon Human2Humanoid, we provide an action execution policy that supports Unitree H1 and Unitree G1.
Our deployment module also includes pretrained policy models—one for Unitree G1 and one for Unitree H1—which can be used directly.

If you would like to train your own policy, please refer to our documentation:
👉 Train

Official Human2Humanoid repository:
👉 https://github.com/LeCAR-Lab/human2humanoid

In addition, if you want to experiment with other policy models for action execution, our deployment module also supports Beyondmimic, TWIST, and more.
Once you obtain your final policy model, proceed to the section on Simulation & Real-World Deployment to execute your actions on real robots.

Beyondmimic Policy Training

Beyondmimic requires input motion data in CSV format, so you must first convert the retargeted robot motion data into CSV.

Switch to the H-ACT directory and create the folder data/beyondmimic to store the results.
Then run:

python scripts/pkl_2_csv.py

The converted motion files will appear in the data/beyondmimic directory.

Training instructions can be found in the official documentation: 👉 https://github.com/HybridRobotics/whole_body_tracking?tab=readme-ov-file#policy-training

TWIST Policy Training

Refer to the official TWIST documentation: 👉 https://github.com/YanjieZe/TWIST

Simulation & Real-World Deployment

After training the desired policies, deployment can be carried out with our unified simulation & real-robot deployment framework RoboJuDo.

RoboJuDo supports:

• Sim2sim & sim2real deployment using Beyondmimic, Human2Humanoid, Twist, and more
• Pretrained policy models for quick real-robot deployment
• A unified, clean interface for integrating custom policy models with minimal effort