Reinforcement Learning for Ball Balancing Using a Robot Manipulator

Explore how to use reinforcement learning to solve control tasks in complex dynamic systems such as a redundant robot manipulator. The goal of the task is to design a controller that can balance a ping-pong ball on a flat surface attached to the end effector of the manipulator.

Model-based control theories like model predictive control (MPC) or other methods can solve such tasks by creating mathematical models of the plant. However, it may become difficult to design such controllers when the plant model becomes complex. Model-free reinforcement learning is an alternative in such situations.

This tutorial walks through how to use Reinforcement Learning Toolbox™ to create and train agents that can perform the ball balancing task while being robust to variabilities in the environment. At the end of the tutorial, you will have learned how to create environments, represent agents through neural networks, and train the networks to satisfactory performance.