Desarrollo de un sistema de interacción humano-humanoide mediante el reconocimiento y aprendizaje del lenguaje corporal

Abstract

At present the advances in the different fields that integrate the robotics, have allowed to be promoted in various areas, even outside the environment industrial, constituting new concepts for it, such as the social and cognitive robotics, which in their simplest synthesis seek development of autonomous devices that are companions for beings human beings and can enrich the daily life of people, as well as allow improvement in the quality of life of these. And it is so, that the development of robots humanoids intended for work such as assistance to the sick and people elderly, and attention to the public in various environments such as airports, hotels, shopping centers, etc. It has opened the doors to new fields of research ranging from the design of the same, to the development of techniques interaction with humans.

This project proposes the development of a system that integrates functionalities that complement the communication process between human beings and humanoid robots, improving human-humanoid interaction, through the implementation of machine learning algorithms, specifically networks artificial neurons, arranged to execute the recognition of the state of encouragement of people (Happy, Sad, Angry, Surprised, Reflective and Normal) using non-verbal language expressed with body language; as well as the ability to "teach" the humanoid robot nonverbal language with which this can complement verbal messages and interact in a manner consistent with the mood of human beings, and that this can generate its own body gestures.

The developed system includes a function based on multilayer neural networks (MLP) used to perform the recognition of mood and classification of the same in the established categories, through the expressed body language using with angular values ​​that describe the orientation of the different body joints, obtained using the Kinect Version 2 sensor, developed by Microsoft. It also includes a function with which you can use a conversation between the user and the Pepper robot, using a system of integrated speech recognition with a response generation tool created on the basis of recurrent neural networks (RNN); which, in turn, taking into account the generated response determines a physical behavior suitable with which humanoid robot Pepper, developed by Aldebaran and Softbank Robotics, can complement the verbal message.

A series of movement sequences were developed, using a technique of which the humanoid robot is programmed by imitation, which indicate specific behaviors associated with the different moods that can recognize the system and consistent with the meaning of the conversation. Additionally, a Generative Adverse Network (GAN) model was used, through from which, exploiting its generative functionality, it was possible to create sequences of movement from the sequences generated with the imitation tool, original and different that allow to generate the sensation of naturalness in the execution of the physical behavior reproduced by the robot