ocs2_mpcnet

OCS2 MPC Net

I just install the latest libraries until 2024.8:

• Ubuntu 24.04

• onnxruntime 1.19

• pybind11 2.13.5

• cuda 12.2

• python 12

• torch 2.4.0

• onnx 1.16.2

• numpy 2.1.0

• black 24.8.0

• pyyaml 6.0.2

Nvidia A500 can train ballbot but cannot train legged robot.

1. Install ONNX Runtime

ONNX Runtime is an inferencing and training accelerator. Here, it is used for deploying learned MPC-Net policies in C++ code. To locally install it, do the following:

• Download the onnxruntime release from Github onnxruntime 1.19
• Unzip the tgz file, and:
  • Move the include folder to /usr/local/include/onnxruntime
  • Move the lib/cmake folder to /usr/local/share/cmake
  • Move the lib/pkgconfig folder to /usr/local/lib/pkgconfig
  • Move the lib to /usr/local/lib64

  • sudo mv include/ /usr/local/include/onnxruntime/
    sudo mv lib/cmake/ /usr/local/share
    sudo mv lib/pkgconfig/ /usr/local/lib
    sudo mv lib/ /usr/local/lib64

• check the install

  • export PKG_CONFIG_PATH=/usr/local/lib/pkgconfig
    export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib64
    
    sudo ldconfig
    pkg-config --cflags --libs libonnxruntime

  • The output should be something like:

  • -I/usr/local/include/onnxruntime -L/usr/local/lib64 -lonnxruntime

2. Build and Run pretrained MPC-Net

2.1 Ballbot

cd ~/ocs2_ws
colcon build --packages-up-to ocs2_ballbot_mpcnet

source ~/ocs2_ws/install/setup.bash
ros2 launch ocs2_ballbot_mpcnet ballbot_mpcnet.launch.py

2.2 Legged Robot

• build the package

cd ~/ocs2_ws
colcon build --packages-up-to ocs2_legged_robot_mpcnet

• launch the simulation without raisim

source ~/ocs2_ws/install/setup.bash
ros2 launch ocs2_legged_robot_mpcnet default.launch.py

• launch the simulation with raisim

source ~/ocs2_ws/install/setup.bash
ros2 launch ocs2_legged_robot_mpcnet raisim.launch.py

3. Train MPC-Net

3.1 Create Python Venv

• install python3-venv

sudo apt-get install python3-venv

• create a virtual environment

cd ~
mkdir venvs && cd venvs
python3 -m venv mpcnet

• install the required packages

source ~/venvs/mpcnet/bin/activate
pip install -r ocs2_mpcnet_core/requirements.txt

below command assumes that you are in the venv and sourced ros2 workspace.

3.2 Ballbot

• train the MPC-Net

source ~/ocs2_ws/install/setup.bash
cd ~/ocs2_ws/src/ocs2_ros2/advance\ examples/ocs2_mpcnet/ocs2_ballbot_mpcnet/ocs2_ballbot_mpcnet
python train.py

• check the tensorboard

cd ~/ocs2_ws/src/ocs2_ros2/advance\ examples/ocs2_mpcnet/ocs2_ballbot_mpcnet/ocs2_ballbot_mpcnet
tensorboard --logdir=runs

to run the trained model, copy the .onnx and .pt file in the runs folder to the policy folder.

3.3 Legged Robot

• train the MPC-Net

source ~/ocs2_ws/install/setup.bash
cd ~/ocs2_ws/src/ocs2_ros2/advance\ examples/ocs2_mpcnet/ocs2_legged_robot_mpcnet/ocs2_legged_robot_mpcnet
python train.py

• train the MPC-Net with the raisim environment

source ~/ocs2_ws/install/setup.bash
cd ~/ocs2_ws/src/ocs2_ros2/advance\ examples/ocs2_mpcnet/ocs2_legged_robot_mpcnet/ocs2_legged_robot_mpcnet
python train.py legged_robot_raisim.yaml

• check the tensorboard

cd ~/ocs2_ws/src/ocs2_ros2/advance\ examples/ocs2_mpcnet/ocs2_legged_robot_mpcnet/ocs2_legged_robot_mpcnet
tensorboard --logdir=runs