Balancing Act: Mastering Beam-and-Ball Control with Reinforcement Learning

Tanushree Burman, Jevon Coney, Chris Rogers, Jennifer Cross, Jivko Sinapov
Proceedings of Robotics in Education (RiE) Conference, 2025
Interface of Beam and Ball Website

Abstract

Current AI education concepts often emphasize supervised machine learning algorithms, while reinforcement learning (RL) education remains limited, typically focusing on introductory concepts. To address the need of a more in-depth RL educational experience, we propose two hands-on robotic activities to introduce foundational RL concepts and foster exploration of RL system design principles. These activities include a LEGO car task and a beam-and-ball balance robot simulation. We conducted a pilot study featuring this two-part curriculum with 13- to 16- year old students and analyzed their learning outcomes quantitatively. Our findings indicate that middle and high school students developed an understanding of basic RL concepts as measured by post-activity reflections. We also discuss next steps to enhance the curriculum, including providing a more interactive experience with RL system design.

Images

Image 1
Image 2
Image 3