Abstract:

The purpose of this article is to introduce a circular maze system as a challenging environment to solve, which could be of interest to the robot and reinforcement learning community. Recently, there have been rapid developments in the fields of machine and reinforcement learning, largely due to the success of deep learning approaches. This has also led to increased interest in the area of learning physics based systems. The circular maze environment that we present here is a low-DoF complex system which could be used to investigate many interesting learning problems. We propose some initial results using both model-free and model-based learning approaches to solve the environment with a single and multiple marbles, to demonstrate some of the challenges that this system presents. We hope to opensource the simulation software and hardware design details of the system in the near future.