How People Talk When Teaching a Robot
Elizabeth S. Kim
Dept. of Computer Science
Yale University
51 Prospect St.
New Haven, CT
eliskim@cs.yale.edu
Dan Leyzberg
Dept. of Computer Science
Yale University
51 Prospect St.
New Haven, CT
dan.leyzberg@yale.edu
Katherine M. Tsui
Dept. of Computer Science
Univ. of Massachusetts,
Lowell
1 University Ave.
Lowell, MA
ktsui@cs.uml.edu
Brian Scassellati
Dept. of Computer Science
Yale University
51 Prospect St.
New Haven, CT
scaz@cs.yale.edu
ABSTRACT
We examine affective vocalizations provided by human teach-
ers to robotic learners. In unscripted one-on-one interac-
tions, participants provided vocal input to a robotic dinosaur
as the robot selected toy buildings to knock down. We find
that (1) people vary their vocal input depending on the
learner’s performance history, (2) people do not wait until a
robotic learner completes an action before they provide in-
put and (3) people na¨ıvely and spontaneously use intensely
affective vocalizations. Our findings suggest modifications
may be needed to traditional machine learning models to
better fit observed human tendencies. Our observations of
human behavior contradict the popular assumptions made
by machine learning algorithms (in particular, reinforcement
learning) that the reward function is stationary and path-
independent for social learning interactions.
We also propose an interaction taxonomy that describes
three phases of a human-teacher’s vocalizations: direction,
spoken before an action is taken; guidance, spoken as the
learner communicates an intended action; and feedback, spo-
ken in response to a completed action.
Categories and Subject Descriptors
I.2.9 [Computing Methodologies]: Artificial Intelligence—
Robotics
General Terms
Experimentation, Human Factors
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
HRI’09, March 11–13, 2009, La Jolla, California, USA.
Copyright 2009 ACM 978-1-60558-404-1/09/03 ...$5.00.
Keywords
Human-robot interaction, na¨ıve teaching, affective vocaliza-
tion, affective input, social learning, reinforcement learning
1. INTRODUCTION
As robots enter human environments as our teammates,
assistants, guides, and therapeutic partners, their value will
depend on their ability to adapt to users and their environ-
ments. For adults, verbal communication is both readily
available and powerful, lending itself naturally as a primary
input modality for robot learners. In this study, we exam-
ine how untrained people naturally teach by observing their
vocalizations as they teach a robot a simple task.
While there are many technologies for automatic speech
recognition (the process of transcribing spoken words into
written words), interpreting the meaning of these words is
beyond the current state of the art for all but limited vo-
cabularies or under very restricted environments. Fortu-
nately, for many utterances directed by an instructor to a
learner, the meaning of the words is often mirrored by the
affect being conveyed. In other words, even without know-
ing what is said, how we say it carries much of the speaker’s
intent. Domesticated animals such as dogs learn from their
caretakers’ approving and prohibitive vocalizations, and it
has been suggested that exaggerated affective expressions in
infant-directed speech (Parentese) are an important learning
input for pre-verbal infants. Affective expressions have been
shown to be useful as an input to artificial learners [5,17].
Automatic affect recognition from voice and other modali-
ties has succeeded in approaching or meeting levels of human
agreement in judgments of affect, within closed application
domains [1,4,17–19,23]. Currently these systems are limited
by constraining users to atypical interaction patterns, or by
training to recognize affect only within constrained applica-
tion scenarios. While these studies offer impressive technol-
ogy, they are technology-driven accounts of affective inter-
action; none directly provides evidence for how untrained
users naturally provide input to a computational or robotic
agent. Nevertheless, as these automatic classifiers approach
human recognition of affect, we expect that robotic learners
will be able to use automatic affect recognition to sustain
23