Learning from the Wizard

This project extends the widely used Learning from Demonstration (LfD) paradigm to the domain of freeform social interaction. Social robots are used today for therapy or education. Because of the inherent difficulties of creating autonomous social interaction behavior, a human secretly controls most of these robots in what is called “Wizard-of-Oz” (WoZ) control. We developed a computational and experimental framework that records these WoZ demonstrations, and uses a hierarchical logistic regression model that allows the robot to choose actions autonomously, in the style of the human demonstrator. Because the demonstrations are sourced from Wizard-of-Oz interactions, we refer to our method as Learning from the Wizard (LfW).

37 participants engaged in Wizard-of-Oz interaction with a robot, playing a cooperative Android tablet game centered around color-mixing. Data from this “demonstration” phase was collected and used as input to the learning algorithm, which used this data to learn an autonomous behavioral policy.
In the final stage, 85 participants took part in a randomized experiment. This experiment had three conditions that differed by what the child played with: Tablet-Only, without the robot; WoZ, with the teleoperated robot; and Autonomous, with the robot acting according to its learned model.

The randomized evaluation experiment showed that the autonomous robot was successfully able to successfully learn pedagogically important social behaviors. Overall, the WoZ robots (WoZ-train and WoZ-Experiment) were very similar to the autonomous robot, all of which were markedly different from interaction with a tablet alone. We also found that children interacted similarly with the autonomous and human-controlled robots, that children treated the autonomous robot more like a peer and were more likely to want to play with the autonomous robot again.