Trey Davis
University of Michigan
Faculty Advisor: Prof. Andreea Bobu
Research Supervisor: Nathan Dennler
Department: Aeronautics and Astronautics
Biography
Trey Davis studies how robots can work with creatives. As an artist working primarily
with ceramics and other sculptural media, he uses his robotics major at the University of
Michigan to better understand the animosity his community feels toward AI development. At
Michigan, he works with Dr. Patrícia Alves-Oliveira to create robotic assistants that leverage
creative psychology to scaffold thinking through thoughtful questions. At MIT, he is studying
robot learning using multimodal inputs, with an eye toward applying similar algorithms to
create robots that adapt to highly variable artist preferences. He believes that in understanding
how robots can fit into the creative process, we come closer to understanding how we can
retain our humanity when working with technology. In his PhD, he aims to challenge his
understanding of these questions and prompt his discipline to think more critically about the
proper role of robotics in future society.
Voice Affect Co-Personalization for Robot Preference Learning
Trey Davis1,2, Nathaniel Dennler2 and Andreea Bobu2
1Department of Robotics, University Michigan – Ann Arbor
2Department of Aeronautics and Astronautics, Massachusetts Institute of Technology
In order for robots to work effectively with humans, they need to adapt to a range of human
preferences. Yet, real human feedback is limited and ambiguous, making efficient adaptation
a core challenge for preference learning. The emotional, or affective, component of speech is
an implicit and useful feedback signal for robot learning. Yet, integrating voice affect into an
online robot learning framework remains under-explored. Through the analysis of voice data
in a previous robot learning experiment, we found that voice affect is situational, dependent
on the user and task framing. We further hypothesize that humans adapt their affect throughout
a task to elicit better performance from the robot. Bringing those insights into voice-based
learning systems, we formulate two algorithms that incorporate vocal affect into a collaborative
robot task. We test these algorithms and our hypothesis in a pilot user study. While the existing
understanding of voice affect frames it as a hidden ground truth to be uncovered by robots,
we expect to see more efficient use of vocal feedback by conceptualizing the decoding of this
signal as a collaborative human-robot effort to construct a shared communication channel.
