Page 258 - Designing Sociable Robots

breazeal-79017 book March 18, 2002 14:24

Grand Challenges of Building Sociable Robots 239

Socially situated learning This challenge problem concerns building a robot that can
learn from humans in social scenarios. In chapters 1 and 2, I presented detailed discussions
of the importance and advantages of socially situated learning for robots. The human social
environment is always changing and is unpredictable. There are many social pressures
requiring that a sociable robot learn throughout its lifetime. The robot must continuously
learn about its self as new experiences shape its autobiographical memory. The robot also
must learn continually from and adapt to new experiences that it shares with others to
establish and maintain relationships. New skills and competencies can be acquired from
others, either humans or other robots. This is a critical capability since the human social
environment is too complex and variable to explicitly pre-program the robot with everything
it will ever need to know.
In this book, I have motivated work with Kismet from the fact that humans naturally offer
many different social cues to help others learn, and that a robot could also leverage from
these social interactions to foster its own learning. Other researchers and I are exploring
specific types of social learning, such as learning by imitation, to allow a human (or in
some cases another robot) to transfer skills to a robot learner through direct demonstration
(Schaal, 1997; Billard & Mataric, 2000; Ude et al., 2000; Breazeal & Scassellati, 2002).

Evaluation metrics As the social intelligence of these robots increases, how will we
evaluate them? Certainly, there are many aspects of a sociable robot that can be measured
and quantified objectively, such as its ability to recognize faces, its accuracy of making eye
contact, etc. Other aspects of the robot’s performance, however, are inherently subjective
(albeit quantifiable), such as the readability of its facial expressions, the intelligibility of
its speech, the clarity of its gestures, etc. The evaluation of these subjective aspects of the
design (such as the believability of the robot) varies with the person who interacts with it.
A compelling personality to one person may be flat to another. The assessment of other
attributes may follow demographic trends, showing strong correlations with age, gender,
cultural background, education, and so forth. Establishing a set of evaluation criteria that
unveils these correlations will be important for designing sociable robots that are well-
matched to the people it interacts with.
If at some point in the future the sociability of these kinds of robots appears to rival our
own, then empirical measures of performance may become extremely difficult to define,
if not pointless. How do we empirically measure our ability to empathize with another, or
another’s degree of self-awareness? Ultimately what matters is how we treat them and how
they treat us. What is the measure of a person, biological or synthetic?

Understanding the human in the loop The question of how sociable robots should fit
into society depends on how these technologies impact the people who interact with them.
We must understand the human side of the equation. How will people interact with sociable