Lab member Rebecca Ramnauth defended her dissertation “Building Intelligent Robots for Social Regulation Therapy” today. The presentation slides from her defense can be found here.
Abstract:
Throughout life, we learn the rules of social behavior by observing others, by exposure to diverse social contexts, and, in some cases, through targeted intervention. More than learning the rules and expectations on how to behave, social regulation involves the dynamic, real-time coordination of one’s internal states and outward behaviors to meet those expectations. Regulation becomes challenging when internal states conflict with external demands, such as in moments of heightened emotion, sensory overload, or social ambiguity. In certain contexts (such as isolation during a global pandemic) or for some individuals (such as those with autism), social regulation can be difficult to achieve and even harder to sustain.
This dissertation positions robots as tools to support the learning of social regulation. Robots are embodied platforms and thus offer unique potential for enabling on-demand, physically co-present interactions. Although the field of robotics has traditionally focused on reliability and precision of motion to achieve physical task assistance, a growing body of literature demonstrates that humans often perceive and respond to robots as social entities. Building on this insight, we explored how robots can provide social value and assistance.
To develop such socially assistive robots, we had to overcome significant technical challenges and rethink the prevailing norms in the field. True social learning unfolds over time and requires exposure to novel real-world situations that test the relevance and adaptability of learned strategies. However, much of what we know about human-robot interaction has emerged from experimental studies in controlled laboratory or clinical environments over short timescales and typically focused on interactions between a single robot and a neurotypical adult. For robots to effectively support social regulation learning, they must operate reliably in unstructured, everyday environments; sustain long-term, repeated engagement with users of various cognitive profiles and social needs; adapt to evolving user behavior and progress; and respond in ways that are not only effective, but also socially appropriate and safe. Every component of this requires overcoming significant computational and non-computational challenges.
Across five core studies presented in this dissertation, we describe our design, development, and deployment of robots that achieve this. While establishing feasibility is a necessary first step in ensuring that a robot operates safely, consistently, and acceptably, our work also examines whether these robots yield meaningful therapeutic outcomes. All experiments were conducted outside of laboratory settings, involved interactions spanning several days to a full month, and took place under challenging real-world conditions, including deployments in participants’ homes during the COVID-19 lockdown. Each study was carefully designed to meet the needs of a highly specialized and protected user population. Collectively, these studies demonstrate the value of robots for encouraging a wide range of regulation skills, including attention sharing, turn-taking, conversational reciprocity, resiliency to interruptions, deep breathing, and emotional de-escalation.
This dissertation presents the first robots developed specifically for adults with autism. It includes one of the only robotic studies to demonstrate continuous learning progression linked to clinical measures of therapeutic efficacy. In addition, it includes the first use of foundation models to deliver unscripted and improvised therapy. It also presents the first robot to address behavioral de-escalation in public spaces while remaining agnostic to users’ age or diagnostic profile.
Advisor: Brian Scassellati
Other committee members:
Marynel Vázquez
Tesca Fitzgerald
Frederick Shic (Seattle Children’s Research Institute)