Modeling trust in assistive human-robot interaction: a data-driven approach
File(s)
Author(s)
Lingg, Nico
Type
Thesis
Abstract
In human-robot interaction, particularly in assistive robotics where users rely on autonomous or semi-autonomous systems for physical support, trust is central to safe and effective collaboration. While research on trust in robotics has grown, many studies use simplified scenarios that fail to capture the complexities of real-world interactions. This thesis addresses these gaps by investigating trust through experimental user studies, introducing novel measurement tools, and developing machine learning models tailored to realistic assistive robotics applications.
Through reviewing empirical studies, we establish that existing work often uses constrained methodological approaches missing critical aspects of embodied interaction. Building on these insights, we develop and validate experimental platforms for studying trust in naturalistic settings. Using an autonomous wheelchair system, we show that trust is strongly influenced by robot performance in real navigation tasks: good performance enhances trust and attitudes, while poor performance harms perception and engagement.
To capture dynamic trust changes, we introduce Trusty, a handheld device enabling continuous trust measurement without disrupting interaction flow. In a user study with our wheelchair system, Trusty is validated against established questionnaires and shown to provide reliable real-time trust feedback.
Leveraging continuous trust data alongside video, gaze tracking, and physiological signals, we design a transformer-based model to predict user trust. Fusing these modalities with auxiliary perceptual tasks improves classification accuracy, with eye-gaze data proving particularly valuable.
We also present AMIGA (Assistive Mobile Interactive Grasping Agent), a mobile manipulation platform combining a UR10e arm with a powered wheelchair base for robust, cost-effective assistive tasks. Using AMIGA, we develop Trust-ACT, an imitation learning approach that incorporates trust annotations for policy refinement. Trust-based trajectory selection improves success rates and execution times in block-stacking tasks.
This work advances reliable, user-centered assistive robotics by deepening the understanding of trust and providing practical tools and frameworks.
Through reviewing empirical studies, we establish that existing work often uses constrained methodological approaches missing critical aspects of embodied interaction. Building on these insights, we develop and validate experimental platforms for studying trust in naturalistic settings. Using an autonomous wheelchair system, we show that trust is strongly influenced by robot performance in real navigation tasks: good performance enhances trust and attitudes, while poor performance harms perception and engagement.
To capture dynamic trust changes, we introduce Trusty, a handheld device enabling continuous trust measurement without disrupting interaction flow. In a user study with our wheelchair system, Trusty is validated against established questionnaires and shown to provide reliable real-time trust feedback.
Leveraging continuous trust data alongside video, gaze tracking, and physiological signals, we design a transformer-based model to predict user trust. Fusing these modalities with auxiliary perceptual tasks improves classification accuracy, with eye-gaze data proving particularly valuable.
We also present AMIGA (Assistive Mobile Interactive Grasping Agent), a mobile manipulation platform combining a UR10e arm with a powered wheelchair base for robust, cost-effective assistive tasks. Using AMIGA, we develop Trust-ACT, an imitation learning approach that incorporates trust annotations for policy refinement. Trust-based trajectory selection improves success rates and execution times in block-stacking tasks.
This work advances reliable, user-centered assistive robotics by deepening the understanding of trust and providing practical tools and frameworks.
Version
Open Access
Date Issued
2025-02-13
Date Awarded
01/11/2025
License URL
Advisor
Demiris, Yiannis
Sponsor
UK Research and Innovation
Grant Number
EP/Y028732/1
Publisher Department
Department of Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)