Aprendizaje por refuerzo para manipulación de objetos con actuador robótico

Abstract

This project will implement a reinforcement learning algorithm for the manipulation of objects by a robotic arm, focused on tasks that a domestic service robot will discuss. It will start from the implementation of the robotic platform, which will have the degrees of freedom and the specific actuator for the manipulation of objects found in everyday situations. Next, the reinforcement learning algorithm will be proposed, starting from the state space that best describes the environment, and the action space that allows tasks to be carried out properly. Likewise, methods that will evaluate the state from direct observation of the environment, such as deep neural networks, will be studied. The training will be carried out mostly in a simulation environment, in order not to jeopardize the physical integrity of the robot in the early stages of learning. At the end of the process, if deemed necessary, the training obtained with the real robot will be refined. Next, the result obtained by the agent in the different tasks learned will be evaluated quantitatively. Finally, it will begin with the implementation of this robotic arm in a prototype of a domestic social robot built entirely by the Santo Tomás University, in order to participate in certificates of social and domestic robotics.