Understanding human intention for human-robot interaction
(English) This doctoral thesis delves into the concept of intention in robotics, aiming to establish a comprehensive and practical definition while exploring its technical and social implications. The research begins by addressing a significant gap in the field: the absence of a clear definition of...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/439693 |
| Acceso en línea: | https://hdl.handle.net/2117/439693 https://dx.doi.org/10.5821/dissertation-2117-439693 |
| Access Level: | acceso abierto |
| Palabra clave: | Human-Robot Interaction Intention understanding Intention inference Perception Anticipation Proactivity Collaborative roles Deep Learning 004 - Informàtica Àrees temàtiques de la UPC::Informàtica |
| Sumario: | (English) This doctoral thesis delves into the concept of intention in robotics, aiming to establish a comprehensive and practical definition while exploring its technical and social implications. The research begins by addressing a significant gap in the field: the absence of a clear definition of intention in robotics, where it is often conflated with motion or goal prediction. To address this, the thesis proposes a novel taxonomy of human intention, synthesising insights from psychology and other disciplines. This taxonomy provides a structured framework for understanding intention, categorising it into various types. This foundational work is further expanded through the development of the Perception-Intention-Action (PIA) cycle, a theoretical framework designed to integrate human intention into the decision-making processes of robots. The PIA cycle enhances traditional Perception-Action models by incorporating intention as a core component, enabling robots to exhibit anticipatory and proactive behaviours, thereby improving Human-Robot Interaction (HRI) and Collaboration (HRC). Furthermore, the thesis introduces collaborative roles (Leader, Follower, Collaborative, Neutral, and Adversarial), expanding traditional frameworks and opening new avenues for robotic behaviour programming. The thesis is structured around three primary use cases: collaborative search, collaborative object transportation, and handover tasks. Each case study serves as a practical demonstration of the PIA cycle or its implications, highlighting the importance of both implicit and explicit intention communication between humans and robots. In the collaborative search task, a mobile application was developed to facilitate explicit communication, with experiments showing that users are willing to communicate their intentions to improve team performance. The collaborative transportation task involved the development of a force-based model that integrates human intention inference with explicit communication, alongside the creation of force/velocity predictors to enhance intention inference. The handover task served as a context to define and posteriorly explore the concepts of anticipation and proactivity, demonstrating that both behaviours can enhance HRI, albeit impacting different aspects of the interaction. The thesis makes several technical contributions, including force/velocity predictors for intention inference and a novel generalization of a Deep Learning (DL) architecture for video processing. These tools not only improve the practical application of the PIA cycle but also offer potential benefits for other research areas. Key findings from the research include the identification of a perceptual threshold beyond which further improvements in prediction accuracy become imperceptible to humans. This suggests that once a ``good enough'' level of accuracy is achieved, additional technical enhancements may not yield noticeable benefits. Additionally, the research revealed a preference among users for natural communication systems over technically robust but less intuitive interfaces. These insights indicate a need for a paradigm shift in HRI development, moving away from increasingly complex predictors and inference engines towards systems that accept robotic fallibility and prioritise natural communication methods. This approach fosters more companion-like interactions instead of utility-based interactions, where robots are seen more like partners instead of mere tools. |
|---|