Sobresaliente "Cum Laude" con Mención Internacional
2021
Society is experiencing a series of demographic changes that can result in an unbalance between the active working and non-working age populations. One of the solutions considered to mitigate this problem is the inclusion of robots in multiple sectors, including the service sector. But for this to be a viable solution, among other features, robots need to be able to interact with humans successfully. This thesis seeks to endow a social robot with the abilities required for a natural human-robot interactions. The main objective is to contribute to the body of knowledge on the area of Human-Robot Interaction with a new, platform-independent, modular approach that focuses on giving roboticists the tools required to develop applications that involve interactions with humans. In particular, this thesis focuses on three problems that need to be addressed: (i) modelling interactions between a robot and an user; (ii) endow the robot with the expressive capabilities required for a successful communication; and (iii) endow the robot with a lively appearance.
The approach to dialogue modelling presented in this thesis proposes to model dialogues as a sequence of atomic interaction units, called Communicative Acts, or CAs. They can be parametrized in runtime to achieve different communicative goals, and are endowed with mechanisms oriented to solve some of the uncertainties related to interaction. Two dimensions have been used to identify the required CAs: initiative (the robot or the user), and intention (either retrieve information or to convey
it). These basic CAs can be combined in a hierarchical manner to create more re-usable complex structures. This approach simplifies the creation of new interactions, by allowing developers to focus exclusively on designing the flow of the dialogue, without having to re-implement functionalities that are common to all dialogues (like error handling, for example).
The expressiveness of the robot is based on the use of a library of predefined multimodal gestures, or expressions, modelled as state machines. The module managing the expressiveness receives requests for performing gestures, schedules their execution in order to avoid any possible conflict that might arise, loads them, and ensures that their execution goes without problems. The proposed approach is also able to generate expressions in runtime based on a list of unimodal actions (an utterance, the motion of a limb, etc...). One of the key features of the proposed expressiveness management approach is the integration of a series of modulation techniques that can be used to modify the robot’s expressions in runtime. This would allow the robot to adapt them to the particularities of a given situation (which would also increase the variability of the robot expressiveness), and to display different internal states with the same expressions.
Considering that being recognized as a living being is a requirement for engaging in social encounters, the perception of a social robot as a living entity is a key requirement to foster human-robot interactions. In this dissertation, two approaches have been proposed. The first method generates actions for the different interfaces of the robot at certain intervals. The frequency and intensity of these actions are defined by a signal that represents the pulse of the robot, which can be adapted to the context of the interaction or the internal state of the robot. The second method enhances the robot’s utterance by predicting the appropriate non-verbal expressions that should accompany them, according to the content of the robot’s message, as well as its communicative intention. A deep learning model receives the transcription of the robot’s utterances, predicts which expressions should accompany it, and synchronizes them, so each gesture selected starts at the appropriate time. The model has been developed using a combination of a Long-Short Term Memory network-based encoder and a Conditional Random Field for generating a sequence of gestures that are combined with the robot’s utterance.
All the elements presented above conform the core of a modular Human-Robot Interaction architecture that has been integrated in multiple platforms, and tested under different conditions.
University:
Universidad Carlos III de Madrid