Autonomous Decision-Making based on Biological Adaptive Processes for Intelligent Social Robots

Abstract

The unceasing development of autonomous robots in many different scenarios drives a new revolution to improve our quality of life. Recent advances in human-robot interaction and machine learning extend robots to social scenarios, where these systems pretend to assist humans in diverse tasks. Thus, social robots are nowadays becoming real in many applications like education, healthcare, entertainment, or assistance. Complex environments demand that social robots present adaptive mechanisms to overcome different situations and successfully execute their tasks. Thus, considering the previous ideas, making autonomous and appropriate decisions is essential to exhibit reasonable behaviour and operate well in dynamic scenarios. Decision-making systems provide artificial agents with the capacity of making decisions about how to behave depending on input information from the environment. In the last decades, human decision-making has served researchers as an inspiration to endow robots with similar deliberation. Especially in social robotics, where people expect to interact with machines with human-like capabilities, biologically inspired decisionmaking systems have demonstrated great potential and interest. Thereby, it is expected that these systems will continue providing a solid biological background and improve the naturalness of the human-robot interaction, usability, and the acceptance of social robots in the following years. This thesis presents a decision-making system for social robots acting in healthcare, entertainment, and assistance with autonomous behaviour. The system’s goal is to provide robots with natural and fluid human-robot interaction during the realisation of their tasks. The decision-making system integrates into an already existing software architecture with different modules that manage human-robot interaction, perception, or expressiveness. Inside this architecture, the decision-making system decides which behaviour the robot has to execute after evaluating information received from different modules in the architecture. These modules provide structured data about planned activities, perceptions, and artificial biological processes that evolve with time that are the basis for natural behaviour. The natural behaviour of the robot comes from the evolution of biological variables that emulate biological processes occurring in humans. We also propose a Motivational model, a module that emulates biological processes in humans for generating an artificial physiological and psychological state that influences the robot’s decision-making. These processes emulate the natural biological rhythms of the human organism to produce biologically inspired decisions that improve the naturalness exhibited by the robot during human-robot interactions. The robot’s decisions also depend on what the robot perceives from the environment, planned events listed in the robot’s agenda, and the unique features of the user interacting with the robot. The robot’s decisions depend on many internal and external factors that influence how the robot behaves. Users are the most critical stimuli the robot perceives since they are the cornerstone of interaction. Social robots have to focus on assisting people in their daily tasks, considering that each person has different features and preferences. Thus, a robot devised for social interaction has to adapt its decisions to people that aim at interacting with it. The first step towards adapting to different users is identifying the user it interacts with. Then, it has to gather as much information as possible and personalise the interaction. The information about each user has to be actively updated if necessary since outdated information may lead the user to refuse the robot. Considering these facts, this work tackles the user adaptation in three different ways. • The robot incorporates user profiling methods to continuously gather information from the user using direct and indirect feedback methods. • The robot has a Preference Learning System that predicts and adjusts the user’s preferences to the robot’s activities during the interaction. • An Action-based Learning System grounded on Reinforcement Learning is introduced as the origin of motivated behaviour. The functionalities mentioned above define the inputs received by the decisionmaking system for adapting its behaviour. Our decision-making system has been designed for being integrated into different robotic platforms due to its flexibility and modularity. Finally, we carried out several experiments to evaluate the architecture’s functionalities during real human-robot interaction scenarios. In these experiments, we assessed: • How to endow social robots with adaptive affective mechanisms to overcome interaction limitations. • Active user profiling using face recognition and human-robot interaction. • A Preference Learning System we designed to predict and adapt the user preferences towards the robot’s entertainment activities for adapting the interaction. • A Behaviour-based Reinforcement Learning System that allows the robot to learn the effects of its actions to behave appropriately in each situation. • The biologically inspired robot behaviour using emulated biological processes and how the robot creates social bonds with each user. • The robot’s expressiveness in affect (emotion and mood) and autonomic functions such as heart rate or blinking frequency.