Teaching Continuous Action Space Reinforcement Learning with a Mobile Robot

• Thiago Martins USP
• Larissa Driemeier USP

Resumo

This paper presents a teaching framework that integrates continuous action space reinforcement learning with a mobile robotics platform. Reinforcement learning (RL) enables agents to learn optimal decision-making policies through interaction with their environment, making it a natural fit for robotics applications. The study examines the challenges and solutions for applying RL algorithms, specifically SARSA and Actor-Critic methods, to continuous state and action spaces. These methods are implemented and compared on a differential-drive robot tasked with maintaining a specific orientation and position. A virtual model is utilized to facilitate safe and efficient learning before deploying the algorithms to a physical robot. This study provides practical insights into teaching reinforcement learning concepts and leveraging them for robotics applications.

Palavras-chave: Education, Neural networks, Decision making, Reinforcement learning, Approximation algorithms, Stability analysis, Mobile robots, Function approximation, Robots, Convergence