Install Dependencies

Install Miniconda/Miniforge (Linux + ARM64/aarch64)

conda create -n g1brainco python=3.8
conda activate g1brainco

Install Dependencies in Conda

# For arm motion control
conda install pinocchio -c conda-forge
pip install meshcat transitions

# For building ROS 2 in Conda
pip install rospkg
pip install -U colcon-common-extensions
# Install empy after the two packages above to avoid empy dependency issues
pip install empy==3.3.2 lark-parser

# For vision applications
pip install opencv-python ultralytics mediapipe pyrealsense2

# For trajectory data loading
pip install pyarrow datasets

# Other dependencies
pip install loguru matplotlib

Install CUDA, Torch, Torchversion in Conda

1. Install CUDA and Torch

Check the System Jetson version

dpkg -l | grep nvidia-l4t-core

Refer to the JetPack - Jetson Linux (L4T) Version Mapping

Jetson Version	JetPack Version	Jetson Model
L4T 35.3.1	5.1.1	Jetson AGX Orin Series, Jetson Orin NX Series, Jetson Orin Nano Series, Jetson Xavier NX series, Jetson AGX Xavier Series

Go to PyTorch for Jetson and download the PyTorch wheel that matches the JetPack version.

Upload the wheel package to the Unitree robot

scp torch-2.1.0a0+41361538.nv23.06-cp38-cp38-linux_aarch64.whl unitree@192.168.123.164:/home/unitree/Downloads

Install CUDA and PyTorch

conda activate g1brainco
cd ~/Downloads
pip install torch-2.1.0a0+41361538.nv23.06-cp38-cp38-linux_aarch64.whl

Verify the installation

python3 -c "import torch; print(torch.__version__); print(torch.version.cuda); print(torch.cuda.is_available()); print(torch.cuda.device_count())"

2. Install Torchversion

Torch	Torchversion	Repository	Python
2.1.0	0.16.0	release/0.16	>=3.8, <=3.11

Download the matched Torchversion. Upload it to the robot.

scp -r vision-release-0.16 unitree@192.168.123.164:/home/unitree/Downloads

Install

cd ~/Downloads/vision-release-0.16  
# Activate the Conda environment
conda activate g1brainco 
export BUILD_VERSION=0.16.0
python3 setup.py install