Full text
| |
| Author(s): |
José Pedro Ribeiro Belo
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | São Carlos. |
| Institution: | Universidade de São Paulo (USP). Instituto de Ciências Matemáticas e de Computação (ICMC/SB) |
| Defense date: | 2022-12-15 |
| Examining board members: |
Roseli Aparecida Francelin Romero;
Plinio Thomaz Aquino Junior;
Glauco Augusto de Paula Caurin;
Denis Fernando Wolf
|
| Advisor: | Roseli Aparecida Francelin Romero |
| Abstract | |
Social robotics represents a branch of human-robot interaction dedicated to developing systems to control robots to operate in unstructured environments in the presence of humans. Social robots must interact with humans by understanding social signals and responding appropriately. Most social robots are still pre-programmed, unable to learn and respond with appropriate actions during an interaction with humans. More elaborate methods use body movements, gaze direction, and body language. In this thesis, the socially acceptable signals commonly used during an interaction are considered for the training of a social robot. An intelligent system was developed to make a robot capable of autonomously deciding which behaviors to emit depending on the human emotional state. For this, the first contribution of this work is an architecture called Social Robotics Deep Q-Network (SocialDQN) is proposed to teach social robots to behave and interact appropriately with humans based on social signals, especially in human emotional states. It offers a framework for the use of social signals to control the robots actions and its learning is carried out through Deep Reinforcement Learning (DRL). A second contribution is the SimDRLSR simulator, which is the first simulator to provide a tool to model humans and their behavior through social signals. The development and validation of the SocialDQN network were carried out with the support of this simulator. The results obtained in several tests carried out in a real environment showed that the system learned satisfactorily to maximize rewards and, consequently, the robot behaved in a socially acceptable way. (AU) | |
| FAPESP's process: | 18/25782-5 - Long-term interaction for interactive behavior learning using deep reinforcement learning |
| Grantee: | José Pedro Ribeiro Belo |
| Support Opportunities: | Scholarships in Brazil - Doctorate |