Dance Success Rate: Success rates for learning the dance were generally subpar.
 The most successful teaching session resulted in 8 high confidence actions for
the hokey-pokey dance and 6 high confidence actions, with 1 medium confidence
action, for the freeform dance.  The least successful resulted in nothing being
taught.

We observed that the low success rates of the dance were due to widespread
confusion about what state the Geo was in, what constitutes a transition, and
the confidence threshold for learning (several people treated the Geo as if it
had computer affordances – ie they thought it should remember what to do with
just one reward, even though the instructions said that they should keep
rewarding the Geo until it’s confident enough).

Social Behavior: We observed that 4 out of 5 participants with engagement Geo
tried to “talk” to the Geo. By contrast, none of the 5 participants with
non-engagement Geo tried to talk to the Geo. We even had one person with
engagement Geo dancing along with the Geo - in a mirror action fashion. We
believe that these behaviors validate our hypothesis that people would adopt a
more natural/social mode of interaction with engagement Geo.

Modeling: We noticed that 7 out of 10 participants attempted to use “model
teaching” on the Geo.  This consists of trying to physically manipulate the
Geo’s joints into doing something and then having it “learn” that. Consequently,
we hypothesize that modeling is a very natural way for people to think about
teaching robots like the Geo, and designate it as an area for future research.

Reward Feedback Regularity: 9 out of 10 participants patted the Geo’s butt for
punishment and stroked its neck for reward. Some of these participants adopted
such a mode of reward feedback even when the experimenter demonstrated pats for
both. We believe that this is a consequence of the Geo’s physical form and the
categorization of action feedback into “reward” and “punishment,” which invites
such behavior archetypes.