Research Statement
Guy Hoffman
Physical
practice
and motor learning are crucial capacities for many activities, yet only
partially understood in humans, and mostly ignored in Artificial Intelligence
and robotics research. We use practice to achieve our most important and
impressive behaviors, from walking and handling objects; through collaboration
and communication; to artistic and athletic performance. Decidedly disparate
from traditional notions of learning, which are mostly abstract and
information-driven, practiced skills are unique insofar as they cannot be
achieved by observation alone, and are inherently first-person and embodied.
One cannot learn to ski in the classroom, and dance ensembles cannot perfect
their fluency through written correspondence.
My research goal is
to develop a computational cognitive model of practice and implement it on
robotic and other artificial agents. This could serve to inform psychological
accounts of practice, explaining behavioral and neurological findings. It will
also result in the design of agents that (i) demonstrate a higher level of
efficiency and quality when repeatedly performing a task, and (ii) perform at a
higher level of adaptation and coordination when performing a task with humans.
Achieving this goal could
have a significant impact on the fields of psychology, robotics, human-machine
interfaces, education, medicine, entertainment, athletics, and beyond. By
modeling the way motor skills are acquired, we uncover a key mechanism of human
development and adult behavior. We may improve the methods for learning new
skills; speed up recovery from injuries; find new ways to embellish athletic
and artistic performance; and reach a deeper understanding of the brain-body
interface. By building agents that not only learn, but practice their skills,
we will be able to design systems that tailor their actions to the task at
hand, and fit the particular behavior and rhythm of their human collaborators.
This will allow us to achieve a more natural human-machine interaction, one
that displays the kind of efficient and satisfying perfor
mance that humans are
accustomed to from each other. It will also serve to promote the yet unachieved
goal of long-term human-robot collaboration in our daily environments—be
it homes, offices, schools, hospitals, or workshops.
Methodology
My research program
rests on four principles (see Figure): (a) the study of human practice; (b) the
computational modeling of physical practice; (c) the design of robotic and
virtual agents which implement these computational models, and (d) the
evaluation of these agents through benchmarks and human-subject studies.
As I have done in the
past, this methodology relies on and encourages an interdisciplinary
collaboration with researchers in related fields, such as cognitive and
behavioral psychology.
Past Research
My past research lays
the foundations for the proposed endeavor. During my early doctoral work at the
MIT Media Lab, I have explored the mechanisms underlying human-human and
human-robot collaboration, including non-verbal behavior, turn-taking, gaze,
and the grounding of shared attention.[1] I have
challenged existing notions of machine learning through human-subject studies
in a collaborative setting,[2]
implemented a working cognitive architecture on a humanoid robot displaying
socially appropriate behavior,[3]
and with my colleagues have conducted human subject experiments measuring the
effect of robotic nonverbal behavior on untrained users.[4]
My focus then shifted to
investigation of fluency and the role of anticipation in human-robot teamwork.
I have designed anticipatory action paradigms and tested them in human-subject
studies within a simulated Markov Decision Process.[5] Later, I
formulated a perceptual-symbol-based cognitive architecture for fluent human-robot
interaction. This system includes onsets of many of my larger research goals:
it enables a robot to learn from repetition, to anticipate human activities,
and thus to increasingly coordinate its actions with those of a human partner.[6]
I have implemented and tested this paradigm on two robots, including a robotic
desk lamp, which I designed and constructed for these experiments. My current
research at the Georgia Institute of Technology further explores these notions
in the context of a human-robot musical performance, designing a robot that
practices with a human musician to achieve rehearsed behavior.
Embodied Cognition and Anticipation
Two principles emerge
from my investigation of joint practice: one is that modeling the attainment of
a physical skill relies on an embodied view of cognition. Most current attempts
to model intelligence, and in particular artificial intelligence, are
constrained to notions of abstract symbolic processing, in which data flows
bottom-up from the sensory to the decision-making and motor systems. Instead, I
have been exploring an embodied approach, viewing concept recall and
decision-making as tightly integrated with perception and exploring the
bidirectional (top-down and bottom-up) relationship between sensing, learning,
and acting.
Second, physical
rehearsal is closely tied to the notion of anticipation. Practice results in
perceptual priming to anticipated sensory input, leading to faster and more
fluent action and interaction. I have shown initial results demonstrating how
anticipation can be used in a robot engaged in team practice to enhance both
the efficiency and the fluency of the interaction.6
Research Program
Cognitive
Modeling Looking ahead, I wish to
continue exploring computational models of motor learning. I am specifically
interested in models which relate high-level cognition to perceptual and
motor
experiences, and the transfer of capabilities from "conscious" to
"automatic"
processing. Moreover, I will investigate anticipation, its role in perceptual
processing, and the relationship between physical first-person practice and the
construction of anticipatory cues. Within this framework, the learned connectivity
between perception and action is of major interest. In the context of joint
practice, I will explore the relation between embodiment, anticipation, and nonverbal
behavior.
Benchmarks There is a need for benchmarks for practiced
behaviors. In the past, I have proposed three benchmarks for fluent joint
activity. I would like to contribute to the development of a battery of metrics
in this nascent research area, hopefully enabling a comparative discussion of
cognitive modeling approaches for practice and joint physical activities.
Human-robot
Interaction Models of physical practice
are appropriate to be validated through research in robotics and human-robot
interaction.
I plan to demonstrate my models on physical robots. These robotic agents need
not be complex humanoids. By focusing on behavior and motion, even simple
non-anthropomorphic robots can be used to evaluate coordinated behavior. I hope
to build agents that can practice on their own and jointly with humans. This
could further our capacity to build robots that work efficiently and fluently,
and could also serve as applied support for recently emerging psychological
models of embodied cognition and anticipation.
Comparative
Studies It makes sense to compare
the performance of these models to both traditional machine learning
approaches, and to real human practice. This calls for comparative
theoretical analysis,
and comparative behavioral study of the above-mentioned implementations, using
the newly construed benchmarks. Within this context, human-subject studies are
invaluable, and I regard them as a central part of my proposed research agenda.
Long-term
adaptation Practiced skills also
affect the capacity of robotic and virtual agents to adapt to humans. In the future, we may be able to design
systems that could—through practice—learn new tasks with little initial
information, and fit the particular behavior and rhythm of their human
counterparts. We want to design robots and virtual agents that can play a long-term role in our daily
lives. Whether these are robotic nurses for elderly living in their homes, or
navigation systems in our cars, adaptively practicing systems could alter the
way we interact with personal technology.
Application
to Human Practice
Finally,
if we are able to successfully model physical practice and implement these
models in artificial agents, we could apply our understandings to the embellishment
of human practice.
A long-term goal of this research is the reverse transfer of our findings in
order to improve human physical performance.
Summary
We are at a highly
opportune moment to explore the notion of physical practice through robotic
agents: robot technologies have reached a maturity and commercial availability
to enable rapid prototyping and extensive laboratory and field research; and a
growing array of techniques and findings in cognitive psychology shed new light
on the way humans perfect their performance. I hope to work at this crossroads
to make a meaningful contribution to this endeavor.