Page 253 - Designing Sociable Robots

breazeal-79017 book March 18, 2002 14:24

234 Chapter 13

prohibition, and soothing. The underlying mechanism would actually be similar to the
previous case, as the human is modulating the robot’s affective state by communicating
these intents. Eventually, this could be extended to having the robot recognize positive and
negative facial expressions.
Recognizing success The same mechanisms for recognizing progress could be used to
recognize success. The ability for the caregiver to socially manipulate the robot’s affective
state has interesting implications for teaching the robot novel acts. The robot may not
require an explicit representation of the desired goal nor a fully specified evaluation function
before embarking upon learning the task. Instead, the caregiver could initially serve as the
evaluation function for the robot, issuing praise, prohibition, and encouragement as she tries
to shape the robot’s behavior. It would be interesting if the robot could learn how to associate
different affective states to the learning episode. Eventually, the robot may learn to associate
the desired goal with positive affect—making that goal an explicitly represented goal within
the robot instead of an implicitly represented goal through the social communication of
affect. This kind of scenario could play an important part in socially transferring new goals
from human to robot. Many details need to be worked out, but the kernel of the idea is
intriguing.
Structured learning scenarios Kismet has two strategies for establishing an appropri-
ate learning environment. Both involve regulating the interaction with the human. The
first takes place through the motivation system. The robot uses expressive feedback to in-
dicate to the caregiver when it is either overwhelmed or under-stimulated. In time, this
mechanism has been designed with the intent that homeostatic balance of the drives
corresponds to a learning environment where the robot is slightly challenged but largely
competent. The second form of regulation is turn-taking, which is implemented in the be-
havior system. Turn-taking is a cornerstone of human-style communication and tutelage.
It forms the basis of interactive games and structured learning episodes. In the near future,
these interaction dynamics could play an important role in socially situated learning for
Kismet.

Quality instruction Kismet provides the human with a wide assortment of expressive
feedback through several different expressive channels. Currently, this is used to help entrain
the human to the robot’s level of competence, and to help the human maintain Kismet’s
“well-being” by providing the appropriate kinds of interactions at the appropriate times.
This could also be used to intuitively help the human provide better quality instruction.
Looks of puzzlement, nods or shakes of the head, and other gestures and expressions could
be employed to elicit further assistance or clarification from the caregiver.