Reinforcement Learning: training and deploying a policy to control inverted pendulum with QUBE - Servo2
This demo models show how to design inverted pendulum controller with "QUBE - Servo 2" of Quanser. And they also show the workflow of plant modeling, control design, code generation, verification, and deployment.
- Simscape™, Simscape Electrical™, Simscape Multibody™
- Deep Learning Toolbox™
- Reinforcement Learning Toolbox™
- MATLAB Coder, Simulink Coder, Embedded Coder®
- MATLAB Support Package for Raspberry Pi Hardware
- Simulink Support Package for Raspberry Pi Hardware
- MATLAB Coder Interface for Deep Learning Libraries
- MEX Compiler
Live scripts for Reinforcement Learning have some commands to train in parallel. The commands are invalid by default. If you want to use them, Parallel Computing Toolbox™ is required.
1. PID Control
The plant model can be linearized around the operating point where the pendulum is inverted. A feedback controller is designed to keep the pendulum inverted. On the ather hand, when the pendulum angle is downward, a steady controller is desinged to keep the pendulum right under.
2. Reinforcement Learning
Requirements for invert the QUBE - Serve 2:
- Oscillate the pendulum whicn is steady at .
- Bring up the pendulum around .
- Keep the angle of pendulum at .
- The motor angle does not exceed the . (Hardware Constraints)
In order to realize the control system satisfying above, Combine the feedback controller created in "1. PID Control", "swing up" reinforcement learing, and "mode select" reinforcement learing.
The reason for building this system is that it is difficult to design a function that meets all the requirements with a single Reinforcement Learning controller. The following document explains the details.
For more information about the modeling, refer to the "RL_multi_control_system.slx".
2.1. "swing up" reinforcement learing
Design SAC agent to get the optimal policy which can swing up the pendulum with the reference for the feedback controller.
2.2. "mode select" reinforcement learing
"mode select" reinforcement learing changes the reference for the feedback controller between constant and the output of "swing up" reinforcement learing. PPO agent is used to get the policy for this "mode select" action.
3. Code generation and verification
Extract the trained policy from the agents, and create a model for deploying controller. Then verify the code execution with SIL and PIL before doing experiment.
Connect Raspberry Pi and QUBE - Servo 2, and run the Raspberry Pi with External Mode.
A set of files for past versions can be downloaded from the following link. However, the past files only contain samples created in the old days.
If you have cloned from GitHub, the past version can be obtained by reverting to the corresponding version below.
- Copyright 2021 The MathWorks, Inc.*
Toshinobu Shintai (2023). Reinforcement-Learning-Inverted-Pendulum-with-QUBE-Servo2 (https://github.com/mathworks/Reinforcement-Learning-Inverted-Pendulum-with-QUBE-Servo2/releases/tag/v2.0.3), GitHub. 取得済み .
プラットフォームの互換性Windows macOS Linux
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!Start Hunting!
See release notes for this release on GitHub: https://github.com/mathworks/Reinforcement-Learning-Inverted-Pendulum-with-QUBE-Servo2/releases/tag/v2.0.3
See release notes for this release on GitHub: https://github.com/mathworks/Reinforcement-Learning-Inverted-Pendulum-with-QUBE-Servo2/releases/tag/v2.0.2
See release notes for this release on GitHub: https://github.com/mathworks/Reinforcement-Learning-Inverted-Pendulum-with-QUBE-Servo2/releases/tag/v2.0.1
See release notes for this release on GitHub: https://github.com/mathworks/Reinforcement-Learning-Inverted-Pendulum-with-QUBE-Servo2/releases/tag/v2.0
See release notes for this release on GitHub: https://github.com/mathworks/Reinforcement-Learning-Inverted-Pendulum-with-QUBE-Servo2/releases/tag/v1.0.1