Virtual environment and automated physical rolling maze as experimental platform for deep reinforcement learning

Abstract

In the context of training competent future engineers, we develop platforms that shall help students to build practical competencies by working on challenging tasks for creative and highly motivating applications. Several of these platforms use systems that autonomously learn to master control tasks. Such systems are typically based on deep reinforcement learning (DRL), and related algorithms are frequently demonstrated by agents that learn to play games. In the following, we report on first results related to a platform where AI agents learn to manoeuvre balls through virtual and physical mazes while avoiding dropping into holes.