Read this in other languages: 中文, 日本語.

The purpose of this project is to use the LeRobot open-source framework to train and test data collected from Unitree robots. Therefore, it is necessary to install the LeRobot-related dependencies first. The installation steps are as follows, and you can also refer to the official LeRobot installation guide:

Download our source code:

cd $HOME

git clone git@github.com:unitreerobotics/unitree_IL_lerobot.git

OR

git clone https://github.com/unitreerobotics/unitree_IL_lerobot.git

Create a virtual environment with Python 3.10 and activate it, e.g. with miniconda:

conda create -y -n lerobot python=3.10
conda activate lerobot

Install LeRobot:

cd unitree_IL_lerobot/lerobot
pip install -e .

NOTE: Depending on your platform, If you encounter any build errors during this step you may need to install cmake and build-essential for building some of our dependencies. On linux: sudo apt-get install cmake build-essential

To control the Unitree robot, some dependencies need to be installed. The installation steps are as follows:

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

If you would like to use the dual-arm operation dataset collected with the Unitree G1 that we provide, you can visit UnitreeG1_DualArmGrasping . To download it, you can refer to the following command:

cd $HOME
mkdir lerobot_datasets && cd lerobot_datasets
git clone https://huggingface.co/datasets/unitreerobotics/UnitreeG1_DualArmGrasping

Note: If a download issue occurs, you may need to install git-lfs. The installation command is as follows:

curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.deb.sh | sudo bash
sudo apt-get install git-lfs
git lfs install

The storage directory structure of the downloaded data is as follows:

lerobot_datasets
└── UnitreeG1_DualArmGrasping
    ├── meta_data
    ├── README.md
    ├── train
    └── videos

Use the following command to set the data storage path:

export DATA_DIR="$HOME/lerobot_datasets/"

Friendly Reminder: By using the configuration files we provide, you can skip the modifications below.

• Modify the policy configuration in lerobot/lerobot/configs/policy:
  • Set dataset_repo_id to UnitreeG1_DualArmGrasping (this relates to the storage directory structure).
    • Notes:
      • If training DP using a dataset in ACT format, you can add the diffusion_aloha.yaml configuration file in lerobot/lerobot/configs/policy; reference.
      • If generating a dataset for DP, be mindful of min and max in override_dataset_stats.
• Modify the environment policy in lerobot/lerobot/configs/env:
  • Mainly adjust state_dim and action_dim in the environment configuration to match the required dimensions.

cd unitree_IL_lerobot/lerobot

• Training Diffusion Policy:

python lerobot/scripts/train.py    policy=diffusion_unitree_real_g1    env=unitree_real_g1     dataset_repo_id=UnitreeG1_DualArmGrasping

• Training ACT:

python lerobot/scripts/train.py    policy=act_unitree_real_g1    env=unitree_real_g1     dataset_repo_id=UnitreeG1_DualArmGrasping

In lerobot/lerobot/scripts, add the eval_g1.py script and then run it.

python lerobot/lerobot/scripts/eval_g1.py --pretrained-policy-name-or-path "$HOME/unitree_imitation/lerobot/outputs/train/2024-10-17/19-45-30_real_world_act_default/checkpoints/100000/pretrained_modell"

Note: --pretrained-policy-name-or-path should be modified according to the location of your trained weights. In the eval_g1.py script, the eval_policy function contains a is_single_hand variable that controls whether to use a single hand or both hands. If is_single_hand is set to True, it indicates the use of a single hand. The use_left_hand variable is used to distinguish between the left or right hand when using a single hand. If use_left_hand is True, it signifies the use of the left hand.

Special reminder: If you have modified the LeRobot code, it is recommended to re-enter the lerobot directory and run pip install -e ..

If you are using a Unitree robot to collect your own data and train, you can refer to the following steps for data collection and conversion:

The open-source teleoperation project avp_teleoperate can be used to collect data using the Unitree G1 humanoid robot. For more details, please refer to the avp_teleoperate project.

The data collected using avp_teleoperate is stored in JSON format, with the structure as shown below. To convert the JSON format into the format required by lerobot, please follow the steps below:

The following conversion steps use this data storage path and format as an example. Assuming the collected data is stored in the $HOME/datasets/ directory under the g1_grabcube_double_hand directory, the format is as follows

g1_grabcube_double_hand/        # Task name
│
├── episode_0001                # First trajectory
│    ├──audios/                 # Audio information
│    ├──colors/                 # Image information
│    ├──depths/                 # Depth image information
│    └──data.json               # State and action information
├── episode_0002
├── episode_...
├── episode_xxx

When generating datasets for LeRobot, it is recommended to ensure that the data naming convention, starting from episode_0, is sequential and continuous. You can use the unitree_utils/sort_and_rename_folders tool to sort and rename the data accordingly.

cd unitree_IL_lerobot
python unitree_utils/sort_and_rename_folders.py --data_dir $HOME/datasets/g1_grabcube_double_hand

Notes: If you're using LeRobot in our project, you can skip the following steps and directly perform the conversion.

• Add unitree_json_formats Data Conversion Tool

Add unitree_json_formats.py in lerobot/lerobot/common/datasets/push_dataset_to_hub. This file contains the tool that reads JSON data and converts it into the format required by LeRobot. (If filtering of the data is needed, you can modify this file.)

Notes: The get_images_data function is responsible for processing image data. You may need to modify the image_key based on the ACT or DP configuration file. By default, it uses the ACT strategy for single camera situations with the naming convention color_0 --> top. For dual camera setups, the naming would be color_0 --> top and color_1 --> wrist.

• Import unitree_json_formats

To enable the use of unitree_json_formats for data conversion, you need to modify lerobot/lerobot/scripts/push_dataset_to_hub.py. Add the following line in the get_from_raw_to_lerobot_format_fn function in push_dataset_to_hub.py:

    elif raw_format=="unitree_json":
        from lerobot.common.datasets.push_dataset_to_hub.unitree_json_format import from_raw_to_lerobot_format

cd unitree_IL_lerobot/lerobot
python lerobot/scripts/push_dataset_to_hub.py --raw-dir $HOME/datasets/ --raw-format unitree_json  --push-to-hub 0 --repo-id UnitreeG1_DualArmGrasping --local-dir  $HOME/lerobot_datasets/UnitreeG1_DualArmGrasping --fps 30

Note: The converted data will be stored in the --local-dir, and --repo-id can be filled in according to your needs.

After the conversion, you can refer to the training steps described above for training and testing.

This code builds upon following open-source code-bases. Please visit the URLs to see the respective LICENSES:

1. https://github.com/huggingface/lerobot
2. https://github.com/unitreerobotics/unitree_dds_wrapper