Teaching approach for deep reinforcement learning of robotic strategies

Abstract This paper presents the development of a teaching approach for Reinforcement Learning (RL) for students at the Faculty of Electrical Engineering, University of Ljubljana. The approach is designed to introduce students to the basic concepts, approaches, and algorithms of RL through examples and experiments in both simulation environments and on a real robot. The approach includes practical programs written in Python and presents various RL algorithms. The Q‐learning algorithm is introduced and a deep Q network is implemented to introduce the use of neural networks in deep RL. The software is user‐friendly and allows easy modification of learning parameters, reward functions, and algorithms. The approach was tested successfully on a Franka Emika Panda robot, where the robot manipulator learned to move to a randomly generated target position, shoot a real ball into the goal, and push various objects into target position. The goal of the presented teaching approach is to serve as a study aid for future generations of students of robotics to help them better understand the basic concepts of RL and apply them to a wide variety of problems.