Sistema de control de un brazo robótico industrial basado en Reinforcement Learning aplicado a manipulación de objetos.

Abstract

Abstract: This Master's thesis analyzes technological improvements in the industrial sector through the use of robotic arms to optimize efficiency and safety in production lines. The main objective was to develop a control system for a robotic arm based on Reinforcement Learning, applied to object manipulation in the CoppeliaSim simulation environment. To achieve this, tools such as GYM and Stable-Baselines3 were implemented, using the Proximal Policy Optimization (PPO) algorithm to enable the robotic arm to interact and learn autonomously, optimizing the execution of pick and place tasks in dynamic environments with changing objects. The results indicate that the model improved its performance over 49 iterations, successfully stabilizing its learning policy. Metrics such as total loss and KL divergence showed consistent progress, demonstrating that the system optimized decision-making and reduced uncertainty in reward prediction. This approach contributes to the improvement of industrial robot control, minimizing the need for manual reprogramming and increasing autonomy in production environments.