macstepien / furuta_pendulum

LQR, MPC and DRL approaches to control the Furuta pendulum.

Furuta Pendulum

LQR, MPC and RL approaches to control the Furuta Pendulum.

Packages in this repository can be divided into two subgroups:

1. Different approaches to solving the control problem:

• furuta_pendulum_de - uses manually derived Differential Equations (hence the name) with manually calculated linearization, which is then used for calculating gains for the LQR controller with swing-up
• furuta_pendulum_drake - in this package the same LQR with a swing-up approach is used, but this time manually derived equations are replaced. Instead Drake library is used to load the URDF of the pendulum, and solely based on it, linearization is calculated, which is once again used to calculate K gain.
• furuta_pendulum_control_toolbox - [this package is deprecated (currently ignored during build), for the MPC OCS2 package should be used] in this approach different library is used - Control Toolbox. It also automatically derives a simulation of the pendulum based on this model (which is in the form of code generated based on the URDF model). There are two approaches used here, one is the same, as in previous ones - LQR with swing-up, but this time Control Toolbox is used to get K. Another approach uses MPC to get an optimal solution.
• furuta_pendulum_rl - in this package Reinforcement Learning approach is used. The pendulum is modeled in the Mujoco, which then is used to create a Gym environment. After training the agent, the actor's network is exported to the final controller, which is implemented in C++.
• furuta_pendulum_ocs2 - Optimal Control for Switched Systems integration that provides MPC.
• furuta_pendulum_ocs2_ros - ROS 2 integration for the OCS2 controller.

2. General utility packages:

• furuta_pendulum_bringup - general launch files, which combine controllers from different packages, with different simulation environments.
• furuta_pendulum_description - URDF of the pendulum and config file with physical properties
• furuta_pendulum_gazebo - simulation config in Gazebo, which uses ros2_control with torque control
• furuta_pendulum_hardware - ROS 2 control configuration for ODrive used in the physical pendulum
• furuta_pendulum_model - data and scripts used for system identification

Building

Some of the source dependencies are specified in the *.repos files, you can use them by calling vcs import src < src/furuta_pendulum/furuta_pendulum.repos in the src directory (depending on your build you should additionally import simulation or hardware dependencies). Some packages require installing additional dependencies if they can't be installed easily with rosdep. I usually add them to the devcontainer, but I didn't have time to automate all of them, so you may have to do it manually, for details refer to the README file of each package.

colcon build --cmake-args -DCMAKE_BUILD_TYPE=Release --symlink-install

Usage

To run the pendulum simulation use one of the launch files from furuta_pendulum_bringup, launch files have the naming convention:

simulator_type_with_controller_type.launch.py

Available simulator_types:

• gazebo
• de_simulation
• hardware

Available controllers:

• drake_controller
• ocs2_controller
• rl_controller

Note

drake_controller uses LQR with energy-based swing-up approach, and parameters are optimized for hardware, thus simulations don't work correctly. ocs2_controller on the other hand doesn't work on hardware, and isn't always reliable in simulations.