TrajBooster: Boosting Humanoid Whole-Body Manipulation via Trajectory-Centric Learning (Deployment)

Jiacheng Liu*, Pengxiang Ding*, Qihang Zhou, Yuxuan Wu, Da Huang, Zimian Peng, Wei Xiao, Weinan Zhang, Lixin Yang, Cewu Lu† , Donglin Wang†
* Equal Controlbution, † Corresponding Authors
Zhejiang University, Westlake University, Shanghai Jiao Tong University & Shanghai Innovation Institute

📋 Contents

• 📋 Contents
• 🏠 Description
• 📚 Usage
  • Unitree G1
  • Deployment
• 🔗 Citation
• 📄 License
• 👏 Acknowledgements

🏠 Description

This repository is an official implementation of the deployment of "TrajBooster: Boosting Humanoid Whole-Body Manipulation via Trajectory-Centric Learning". It requires an Unitree G1 with Dex-3 hands and a personal computer. All communications between the robot and the PC are via Wi-Fi. Our code is based on Walk-These-Ways and Unitree SDK2. Our hardware system is also open-sourced, you can first refer to it to reimplement the teleoperation system. Once you successfully build a system, you can follow the instructions in the Usage. For simple usage, we just provide an example checkpoint named deploy.onnx.

📚 Usage

Unitree G1

NOTE: It is recommended that you connect a screen, a keyboard, and a mouse to the Unitree G1 to use the board on it as a computer.

First of all, you should install PyTorch on the Nvidia Jetson Orin of the Unitree G1, which is different from the process on a normal Ubuntu PC. For this step, please refer to the official instruction.

Then, you can use any ways to put this code to the Orin and enter the directory. Install the required packages by running:

pip install -r requirements.txt

For controlling robot, we use the cpp-based Unitree SDK2, which means you should first compile the g1_control.cpp for Unitree G1 and hand_control.cpp for Dex-3. We have prepared the required CMakeLists.txt for you, so you only need to run the following command lines:

cd unitree_sdk2
rm -rf build
mkdir build && cd build
cmake ..
make

Then the runnable binary files will be in unitree_sdk2/build/bin. You also need to install the g1_gym_deploy by running:

cd g1_gym_deploy && pip install -e .

Deployment

🎯 Before deployment, please run L1+A L2+R2 L2+A L2+B to close G1's initial control process, if successful, you will see the robot hang up its arm after L2+A and lose efforts after L2+B.

For TCP communications, you should determine the IP address of your PC and robot by running:

ifconfig | grep inet

Set the IP addresses in the code to the correct value.

A. Run the robot control program on robot (robot terminal 2):

cd unitree_sdk2/build/bin && ./g1_control eth0 (or eth1)

B. Run the inference thread to make RL policy control robot on robot (robot terminal 3):

Please select one of the following control modes:

1. For Teleoperation Mode:

python g1_gym_deploy/scripts/deploy_policy.py 

2. For Autonomous Control Mode:

python g1_gym_deploy/scripts/deploy_policy_infer.py

C. After putting the robot on the ground, push the R2 button of the joysticker, make the robot stand on the ground, and push R2 again.

NOTE: We strongly recommend you to really deploy the system after you really understand function of all files, otherwise there can be some troubles.

🔗 Citation

If you find our work helpful, please cite:

@article{liu2025trajbooster,
  title={TrajBooster: Boosting Humanoid Whole-Body Manipulation via Trajectory-Centric Learning},
  author={Liu, Jiacheng and Ding, Pengxiang and Zhou, Qihang and Wu, Yuxuan and Huang, Da and Peng, Zimian and Xiao, Wei and Zhang, Weinan and Yang, Lixin and Lu, Cewu and Wang, Donglin},
  journal={arXiv preprint arXiv:2509.11839},
  year={2025}
}

📄 License

This project is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License .

👏 Acknowledgements

• Walk-These-Ways: Our robot deployment code is based on walk-these-ways.
• Unitree SDK2: We use Unitree SDK2 library to control the robot.
• OpenHomie: This work is developed based on OpenHomie codebase.