​RoboMimic Deploy​​ is a multi-policy robot deployment framework based on a state-switching mechanism. Currently, the included policies are designed for the ​​Unitree G1 robot (29-DoF)​​.

It is recommended to run training or deployment programs in a virtual environment. We suggest using Conda to create one.

Use the following command to create a virtual environment:

conda create -n robomimic python=3.8

conda activate robomimic

PyTorch is a neural network computation framework used for model training and inference. Install it with the following command:

conda install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 pytorch-cuda=12.1 -c pytorch -c nvidia

Clone the repository via git:

git clone https://github.com/ccrpRepo/RoboMimic_Deploy.git

Navigate to the directory and install:

cd RoboMimic_Deploy
pip install numpy==1.20.0
pip install onnx onnxruntime

git clone https://github.com/unitreerobotics/unitree_sdk2_python.git
cd unitree_sdk2_python
pip install -e .

python deploy_mujoco/deploy_mujoco.py

1. Connect an Xbox controller.
2. Run the simulation program:

python deploy_mujoco/deploy_mujoco.py

3. Press the ​​Start​​ button to enter position control mode.

4. Hold ​​R1 + A​​ to enter ​​LocoMode​​, then press BACKSPACE in the simulation to make the robot stand. Afterward, use the joystick to control walking.

5. Hold ​​R1 + X​​ to enter ​​Dance​​ mode—the robot will perform the Charleston. In this mode:

   • Press ​​Select​​ at any time to switch to damping protection mode.
   • Hold ​​R1 + A​​ to return to walking mode (not recommended).
   • Press ​​Start​​ to return to position control mode.

6. The terminal will display a progress bar for the dance. After completion, press ​​R1 + A​​ to return to normal walking mode.

7. In ​​LocoMode​​, pressing ​​R1 + Y​​ triggers a Martial arts movement —​ ​use only in simulation​​.

8. In ​​LocoMode​​, pressing ​​L1 + Y​​ triggers a Martial arts movement(Failed) —​ ​use only in simulation​​.

9. In ​​LocoMode​​, pressing ​​R1 + B​ triggers a Kick movement(Failed) —​ ​use only in simulation​​.

1. Power on the robot and suspend it (e.g., with a harness).

2. Run the deploy_real program:

python deploy_real/deploy_real.py

3. Press the ​​Start​​ button to enter position control mode.
4. Subsequent operations are the same as in simulation.

The current framework does not natively support deployment on G1 robots equipped with Orin NX platforms. Preliminary analysis suggests compatibility issues with the unitree_python_sdk on Orin systems. For onboard Orin deployment, we recommend the following alternative solution:

• Replace with unitree_sdk2 (official C++ SDK)
• Implement a dual-node ROS architecture:
  • C++ Node: Handles data transmission between robot and controller
  • Python Node: Dedicated to policy inference

The Mimic policy does not guarantee 100% success rate, particularly on slippery/sandy surfaces. In case of robot instability:

• Press F1 to activate PassiveMode (damping protection)
• Press Select to immediately terminate the control program

Currently the only verified stable policy on physical robots:

⚠️ Important Precautions:

• Palm Removal Recommended: The original training didn't account for palm collisions (author's G1 lacked palms)
• Initial/Final Stabilization: Brief manual stabilization may be required when starting/ending the dance
• Post-Dance Transition: While switching to Locomotion/PositionControl/PassiveMode is possible, we recommend:
  • First transition to PositionControl or PassiveMode
  • Provide manual stabilization during transition

All other movements are currently not recommended for physical robot deployment.

Always master operations in simulation before attempting physical robot deployment.