DataDrivenMPC

Model predictive control based on data-driven model

Install

Requirements

• Compiler supporting C++17
• Tested on Ubuntu 20.04 / ROS Noetic and Ubuntu 18.04 / ROS Melodic

Dependencies

This package depends on

• libtorch (PyTorch C++ Frontend)
• NMPC

Some tests depend on

• PyBullet

Installation procedure

It is assumed that ROS is installed.

1. Follow the official instructions to download and extract the zip file of libtorch.
2. Setup catkin workspace.

   $ mkdir -p ~/ros/ws_ddmpc/src
   $ cd ~/ros/ws_ddmpc
   $ wstool init src
   $ wstool set -t src isri-aist/NMPC git@github.com:isri-aist/NMPC.git --git -y
   $ wstool set -t src isri-aist/DataDrivenMPC git@github.com:isri-aist/DataDrivenMPC.git --git -y
   $ wstool update -t src

3. Install dependent packages.

   $ source /opt/ros/${ROS_DISTRO}/setup.bash
   $ rosdep install -y -r --from-paths src --ignore-src

4. Build a package.

   $ catkin build data_driven_mpc -DCMAKE_BUILD_TYPE=RelWithDebInfo -DLIBTORCH_PATH=<absolute path to libtorch> --catkin-make-args all tests

<absolute path to libtorch> is the path to the directory named libtorch that was extracted in step 1.

Examples

Make sure that it is built with --catkin-make-args tests option.

MPC for Van der Pol oscillator

Control the Van der Pol oscillator by the learned state equation.

$ rosrun data_driven_mpc TestMpcOscillator

MPC for walking with pushing

Control the CoM motion of robot and object by combining the known CoM-ZMP model and the learned object dynamics model.

$ rosrun data_driven_mpc TestMpcPushWalk --gtest_filter=*.RunMPC
$ rosrun data_driven_mpc plotTestMpcPushWalk.py

MPC for cart pushing

Control the position and angle of the one wheel cart on the PyBullet dynamics simulator.

# 3-second simulation
$ rostest data_driven_mpc TestMpcCart.test enable_gui:=true --text
# Endless simulation
$ rostest data_driven_mpc TestMpcCart.test no_exit:=true enable_gui:=true --text

MPC for walking with cart pushing

Control the robot CoM and the position and angle of the one wheel cart on the PyBullet dynamics simulator. The robot CoM-ZMP model is known. The object dynamics model is learned.

$ rostest data_driven_mpc TestMpcCartWalk.test enable_gui:=true --text