Teaching Reinforcement Learning using a Physical Robot

摘要

This paper presents a little crawling robot as a didactic instrument for teaching reinforcement learning. The robot learns a forwardwalking policy from scratch in less than 20 seconds of reinforced sensorimotor interactions. The state space consists of two discretized dimensions, where the behavior is visualizable and comprehensible. In laboratory tutorials, students conduct experiments with a Toolbox written in Java, enabling to vary signicant parameters such as the exploration rate or the discounting factor. Our experience as well students feedback in using the crawler reects: