Reinforcement Learning in Educational Robotics: A Framework for Human-Robot Interaction in 21st Century Classrooms

Abstract

Integrating reinforcement learning (RL) into educational robotics is not merely a trend; it subtly transforms the interaction between children and machines within the classroom environment. This project utilises a modular reinforcement learning engine integrated with the established Gazebo-ROS-OpenAI Gym framework, designed to instruct a robot on which student groups warrant a greeting and which do not. Two algorithms, PPO and DQN, were executed for ten thousand episodes within a pixelated simulation of a lecture hall, enabling the robot to navigate around vacant desks while advancing towards the most dynamic clusters. Numerical data does not convey the complete narrative, although it remains striking: Reinforcement learning outperformed hand-coded rules by 42.3 percent in engagement scores and remarkably reduced crashes by 93.8 percent when seats were closely arranged. The planner achieved a nearly precise increase of 38.6 percent for the PPO configuration, and the policies ceased seeking superior rewards after approximately 6,000 iterations. The tallies, as unreliable as laboratory measurements typically are, suggest that real-time learning enhances a robot’s sensibility, improves its timing, and imparts a degree of social finesse when it enters a classroom of pupils. Ultimately, the wiring may be replicated across many floor designs without necessitating a mental reconfiguration, rendering the kit suitable for diverse applications, from a serene study area to a chaotic science fair.