Experimental platform for deep reinforcement learning using 3D simulation and a physical demonstrator of rolling mazes

Abstract

A primary objective in the education of engineers is to develop students to be valuable employees for their future companies or competent founders of own businesses. To meet the objective students must build up and strengthen skills and competencies required for their future profession. There is nothing wrong, though, with doing so by working on creative or even playful tasks. A focus on educational aspects is a chance to offer students challenging and highly motivating tasks, while our personal experience shows that high intrinsic motivation and creative working environments typically result in very good learning effects.

In this context, we create systems or "platforms" that serve students as working environments for their bachelor's and master's theses. These platforms should combine technologies and methods from different fields such as deep learning, image processing, software development, electronics, and mechatronics, to name a few. This puts us in the position to offer tasks from various technological fields, and it requires students to take interdisciplinary aspects into account.

The system subject to this paper is motivated by a famous traditional maze game. It consists of a wooden box which contains a board with start location, walls, holes, and finish. Players use two knobs to tilt the board in x and y direction, respectively, and must maneuver a ball along a given path from start to finish. A player loses when the ball drops into a hole and wins when the ball reaches the finish.

The overall objective of our activities is to develop deep reinforcement agents, being a specific branch of artificial intelligence, that learn to play the game in a software simulation as well as by controlling the original physical maze. In this paper we report on intermediate results on modelling and solving the game in a software environment and on the analysis of the physical game by image processing methods.