This is our robot Junior. We are studying how the presence of a teacher affects Junior’s learning of affordances. We do this by comparing self-exploration based affordance learning with teacher controlled learning. We try to identify the ways teachers facilitate learning of affordances so that we can improve Junior’s learning and interaction methods.

Read more about Junior.

Read more about affordance learning.

Read more about our experiments.

Read more about our findings.

People are convinced that hard-coding everything into a robot is impossible and that they should be able to learn things. However the way that machine/robot learning has been studied assumes that the “world experience” needed to learn something can systematically be acquired or can spontaneously arise in random exploration. We are inspired by the way parents structure the surrounding of their infants such that important learning opportunities are made available for the developing infant. We want to explore this ability of human by making them structure the workspace of a robot learning affordances of objects.

We believe that non-social and social learning are complementary and are both necessary. We find that systematic training is useful because the designer of the robot can identify better than a common person all the possible things that the robot needs to learn. However we find that most of the effort in this type of training goes to learning that “nothing happens” in most situations. Unlike the systematic case, in the social case almost always something happens. People configure the environment such that an interesting interaction will occur, and they make happen the things they think are “successful” goals of actions. In this sense they provide positive examples much more efficiently than a systematic or random training process. Finally, the random case lets the robot play by itself, which can sometimes lead to unique opportunities of practicing actions by repeating them over and over (see video below).

This study was conducted as a course project in the “Human Robot Interaction” class at the College of Computing, Georgia Tech.