Navigation and human-robot interaction using reinforcement learning

Advances in artificial intelligence (AI) have revolutionized human-robot interaction and paved the way for increasingly sophisticated autonomous systems. Reinforcement learning (RL), a subfield of machine learning, has emerged as a powerful paradigm for training intelligent agents through trial and error. In the field of human-robot interaction, RL offers an unprecedented opportunity to give robots the ability to perceive and control social situations. This thesis presents a simulation and an architecture designed to facilitate the learning of social affordances by an autonomous agent, enabling seamless interactions with humans. We propose a framework that combines RL algorithms with a comprehensive understanding of social affordances, with the primary goal to develop an agent capable of interacting with a human, grasping his attention and carrying him around specifics goal areas. To evaluate the efficacy of our architecture, simulation trials were conducted using simulated humans with different characteristics, to assess the agent's ability to learn in different situations. In addition, an attempt was made to initialise a real-world experiment using Turtlebot robot, which can serve as a versatile platform for testing the agent acquired knowledge in a table scenario, focusing only in a navigation problem. The simulation results clearly demonstrated the success of the designed learning architecture in integrating the knowledge acquired by the agent across different tasks, resulting in a great performance in terms of accumulated reward. Furthermore, in the experimental setup, the navigation problem tackled with the Turtlebot yielded excellent results. This thesis contributes to the burgeoning field of human-robot interactions by leveraging RL techniques to enable robots to comprehend and engage with social affordances.