Mon, Oct 29, 2018 @ 01:00 PM - 03:00 PM
Title: Human-Robot Learning: Computational Personalization For Socially Assistive Robotics
Time: 01:00 PM on Monday, October 29th, 2018
Location: RTH 406
Ph.D. Candidate: Caitlyn Clabaugh
Prof. Maja Matarić
Prof. Gaurav Sukhatme
Prof. Gisele Ragusa
Socially assistive robotics (SAR) seeks to support human care and development through long-term, socially co-present interaction, supplementing the efforts of clinicians and educators. One primary objective in SAR is the personalization or tailoring of interaction to meet the unique and evolving abilities, preferences, and needs of individuals. This dissertation proposes a theoretical framework for human-robot learning (HRL) to enable computational personalization.
HRL is formalized as a hierarchical decision-making problem, wherein the robot selects actions to maximize an individual's psychosocial state and progress toward some assistive goal. The framework groups SAR actions into abstract categories based on theories from psychology and linguistics. Actions within each category are selected by a local policy or controller. The controllers themselves are activated by a meta-controller, an overarching heuristic or algorithm that controls the flow of the intervention. In this way, the framework hierarchically decomposes the large state-action spaces of SAR into more tractable subspaces for computational personalization.
The proposed framework was instantiated as an individualized SAR intervention for early childhood math. To validate the framework and its instantiation, a short-term study was conducted with typically developing children in a general preschool classroom. The data collected informed iterative design and computational personalization, culminating in a long-term, in-home SAR intervention for children with autism spectrum disorder (ASD). In this context, individualization was framed as a reinforcement learning problem, adapting the SAR's instruction and feedback to each child over many interactions.
The fully autonomous SAR system was deployed for month-long interventions in the homes of ten children with ASD. The single-subject study found that the SAR system successfully individualized its instruction and feedback to each child participant over time. Additionally, all participants showed improvement in their mathematics skills and long-term retention of intervention content, demonstrating the quality of the individualization. This research designed, developed, and deployed a novel, fully autonomous, long-term, in-home, individualized SAR intervention for children with diverse needs. As a broader contribution, this dissertation formalizes the problem of HRL and offers a validated, theoretical framework to inform future research at the intersection of artificial intelligence and human-machine interaction.
Audiences: Everyone Is Invited
Contact: Lizsl De Leon