Unitree_RL

中文文档

Quadruped robot reinforcement learning deployment method based on the unitree_guide state machine. The model trained through Isaac gym is successfully simulated in gazebo and deployed on the real robot. By modifying CMakeLists.txt, the switch between simulation and deployment can be achieved.

Considering the file size, only go1_description is reserved in src/unitree_ros/robots. If you need to use other robots, please add the urdf file yourself.

A trained weight file for Unitree Go1 walking on flat ground is provided.

If this repository is helpful to you, please help click stars, thank you.

Prepare

Download libtorch: libtorch-cxx11-abi-shared-with-deps-1.10.0+cu113.zip

https://download.pytorch.org/libtorch/cu113/libtorch-cxx11-abi-shared-with-deps-1.10.0%2Bcu113.zip

Remember the folder path after decompression: /yourpath/libtorch

Download the repository:

git clone https://github.com/dstx123/unitree_rl.git

Compile

libtorch path

Modify the libtorch path in the file unitree_guide/unitree_guide/CMakeLists.txt:

set(Torch_DIR /yourpath/libtorch/share/cmake/Torch)

Compile the project

Run in the directory where src is located

catkin_make

Load custom model

If you need to load a model you trained yourself, you may need to make the following adjustments to the code:

Weight model

Modify the model path in the file unitree_guide/unitree_guide/CMakeLists.txt:

set(BODY_MODEL_PATH "${CMAKE_CURRENT_SOURCE_DIR}/model/body.jit")
set(ADAPT_MODEL_PATH "${CMAKE_CURRENT_SOURCE_DIR}/model/adapt.jit")
add_definitions(-DBODY_MODEL_PATH="${BODY_MODEL_PATH}")
add_definitions(-DADAPT_MODEL_PATH="${ADAPT_MODEL_PATH}")

Here you can customize and add your own models.

Modify the corresponding model loading method in unitree_guide\unitree_guide\src\FSM\State_RL.cpp. Please pay attention to the function void State_RL::_loadPolicy() and void State_RL::_action_compute()

Input and output variables

Please check the order of input and output variables and related parameters. Mainly seen in unitree_guide\unitree_guide\src\FSM\State_RL.cpp and unitree_guide\unitree_guide\include\FSM\State_RL.h

Running

Simulation

start gazebo:

source devel/setup.bash
roslaunch unitree_guide gazeboSim.launch

Create a new terminal and run the controller:

source devel/setup.bash
sudo ./devel/lib/unitree_guide/junior_ctrl

Press the key 2 and the robot will stand up. Then press the key 4, the robot will enter the reinforcement learning state machine. After the robot stands firm, control the movement through the keyboard.

In the terminal, you can use the 'w' 's' 'a' 'd' keys to control the robot's x and y-axis speed, and the 'j' 'l' keys to control the robot's rotation.

Physical deployment

Make the following changes in the file unitree_guide/unitree_guide/CMakeLists.txt:

set(SIMULATION OFF)         # Use Gazebo or not, ON or OFF
set(REAL_ROBOT ON)          # Link real robot or not, ON or OFF

Recompile:

catkin_make

Connect the computer and the robot through a network cable. Please make sure you can ping the robot. Please refer to Unitree repository for details.

Start the robot, and after the robot stands up, run this in the terminal:

sudo ./devel/lib/unitree_guide/junior_ctrl

At this time, the robot will sit down and enter the passive state. Press L2+A, the robot will stands up. Then press 'start'. At this time, the robot enters the reinforcement learning state machine, controlling the robot's x- and y-axis speeds through the left joystick, and controlling the robot's rotation through the right joystick.

Acknowledgments

Thanks to Unitree official for providing the rl_ws project.