Aplicación del algoritmo de aprendizaje por refuerzo state-action-reward-state-action (SARSA) para la generación de trayectorias óptimas en plataformas robóticas

Abstract

Path planning for mobile robots is crucial to improve efficiency and reduce operational risks in the 4.0 industry. Companies are looking to implement advanced technologies such as artificial intelligence and reinforcement learning to automate industrial processes. This study was developed using the Kanban methodology, allowing the work to be broken down into several stages and tasks for monitoring and control. The implementation of the reinforcement learning algorithm SARSA to generate optimal trajectories in mobile robotic platforms is addressed, focusing on adapting and applying this "on-policy" algorithm, which updates its action values based on the direct experience with the environment. The experimental process included implementing the SARSA algorithm in a simulated environment for the autonomous-capable KUKA YouBot robot and a LIDAR sensor on an Nvidia Jetson AGX Orin module. The agent interacted with the environment through training episodes, learning through ε-greedy policy exploration and exploitation of the current and next actions available and computed based on the current and next states, respectively. The trained models were tested in a real environment with the KUKA YouBot robot to validate their performance under practical conditions. Finally, the results were integrated with a larger project using Deep Learning techniques to optimize autonomous trajectories in mobile robots within industrial processes, demonstrating the feasibility and advantages of using reinforcement learning algorithms in advanced robotic applications.